C++ Coding Guideline
Follow Google style for C++, with two exceptions:
Each line of text may contain up to 100 characters.
The use of C++ exceptions is allowed.
Use C++17 features such as smart pointers, braced initializers, lambda functions, and
Use Doxygen to document all the interface code.
We have some comments around symbols imported by headers, some of those are hinted by include-what-you-use. It’s not required.
We use clang-tidy and clang-format. You can check their configuration in the root directory of the XGBoost source tree.
We have a series of automatic checks to ensure that all of our codebase complies with the Google style. Before submitting your pull request, you are encouraged to run the style checks on your machine. See R Coding Guideline.
Python Coding Guideline
R Coding Guideline
We follow Google’s C++ Style guide for C++ code.
This is mainly to be consistent with the rest of the project.
Another reason is we will be able to check style automatically with a linter.
When needed, you can disable the linter warning of certain line with
We use roxygen for documenting the R package.
The following steps are followed to add a new Rmarkdown vignettes:
Add the original rmarkdown to
doc/R-package/Makefileto add the markdown files to be build.
Clone the dmlc/web-data repo to folder
Now type the following command on
This will generate the markdown, as well as the figures in
doc/R-package/index.mdto point to the generated markdown.
Add the generated figure to the
If you already cloned the repo to doc, this means
Create PR for both the markdown and
You can also build the document locally by typing the following command at the
The reason we do this is to avoid exploded repo size due to generated images.
R package versioning
Testing R package with different compilers
You can change the default compiler of R by changing the configuration file in home
directory. For instance, if you want to test XGBoost built with clang++ instead of g++ on
Linux, put the following in your
Be aware that the variable name should match with the name used by
R CMD config CXX17
Registering native routines in R
According to R extension manual,
it is good practice to register native routines and to disable symbol search. When any changes or additions are made to the
C++ interface of the R package, please make corresponding changes in
src/init.c as well.
Generating the Package and Running Tests
The source layout of XGBoost is a bit unusual to normal R packages as XGBoost is primarily written in C++ with multiple language bindings in mind. As a result, some special cares need to be taken to generate a standard R tarball. Most of the tests are being run on CI, and as a result, the best way to see how things work is by looking at the CI configuration files (GitHub action, at the time of writing). There are helper scripts in
R-package/tests/helper_scripts for running various checks including linter and making the standard tarball.
Running Formatting Checks Locally
Once you submit a pull request to dmlc/xgboost, we perform two automatic checks to enforce coding style conventions. To expedite the code review process, you are encouraged to run the checks locally on your machine prior to submitting your pull request.
We use a combination of linters to enforce style convention and find potential errors. Linting is especially useful for scripting languages like Python, as we can catch many errors that would have otherwise occurred at run-time.
For Python scripts, pylint, black and isort are used for providing guidance on coding style, and mypy is required for type checking. For C++, cpplint is used along with
clang-tidy. For R,
lintr is used.
To run checks for Python locally, install the checkers mentioned previously and run:
cd /path/to/xgboost/ python ./tests/ci_build/lint_python.py --fix
To run checks for R:
cd /path/to/xgboost/ Rscript tests/ci_build/lint_r.R $(pwd)
To run checks for cpplint locally:
cd /path/to/xgboost/ python ./tests/ci_build/lint_cpp.py
See next section for clang-tidy. For CMake scripts:
Lastly, the linter for jvm-packages is integrated into the maven build process.
Clang-tidy is an advance linter for C++ code, made by the LLVM team. We use it to conform our C++ codebase to modern C++ practices and conventions.
To run this check locally, run the following command from the top level source tree:
cd /path/to/xgboost/ python3 tests/ci_build/tidy.py
Also, the script accepts two optional integer arguments, namely
--cuda. By default they are both set to 1, meaning that both C++ and CUDA code will be checked. If the CUDA toolkit is not installed on your machine, you’ll encounter an error. To exclude CUDA source from linting, use:
cd /path/to/xgboost/ python3 tests/ci_build/tidy.py --cuda=0
Similarly, if you want to exclude C++ source from linting:
cd /path/to/xgboost/ python3 tests/ci_build/tidy.py --cpp=0
Guide for handling user input data
This is an in-comprehensive guide for handling user input data. XGBoost has wide verity
of native supported data structures, mostly come from higher level language bindings. The
inputs ranges from basic contiguous 1 dimension memory buffer to more sophisticated data
structures like columnar data with validity mask. Raw input data can be used in 2 places,
firstly it’s the construction of various
DMatrix, secondly it’s the in-place
prediction. For plain memory buffer, there’s not much to discuss since it’s just a
pointer with a size. But for general n-dimension array and columnar data, there are many
subtleties. XGBoost has 3 different data structures for handling optionally masked arrays
(tensors), for consuming user inputs
ArrayInterface should be chosen. There are many
existing functions that accept only plain pointer due to legacy reasons (XGBoost started
as a much simpler library and didn’t care about memory usage that much back then). The
ArrayInterface is a in memory representation of
defined by numpy or the
__cuda_array_interface__ defined by numba. Following is a
check list of things to have in mind when accepting related user inputs:
[ ] Is it strided? (identified by the
[ ] If it’s a vector, is it row vector or column vector? (Identified by both
[ ] Is the data type supported? Half type and 128 integer types should be converted before going into XGBoost.
[ ] Does it have higher than 1 dimension? (identified by
[ ] Are some of dimensions trivial? (shape[dim] <= 1)
[ ] Does it have mask? (identified by
[ ] Can the mask be broadcasted? (unsupported at the moment)
[ ] Is it on CUDA memory? (identified by
datafield, and optionally
Most of the checks are handled by the
ArrayInterface during construction, except for
the data type issue since it doesn’t know how to cast such pointers with C builtin types.
But for safety reason one should still try to write related tests for the all items. The
data type issue should be taken care of in language binding for each of the specific data
input. For single-chunk columnar format, it’s just a masked array for each column so it
should be treated uniformly as normal array. For input predictor
X, we have adapters
for each type of input. Some are composition of the others. For instance, CSR matrix has 3
potentially strided arrays for
values. No assumption
should be made to these components (all the check boxes should be considered). Slicing row
of CSR matrix should calculate the offset of each field based on respective strides.
For meta info like labels, which is growing both in size and complexity, we accept only
masked array at the moment (no specialized adapter). One should be careful about the
input data shape. For base margin it can be 2 dim or higher if we have multiple targets in
the future. The getters in
DMatrix returns only 1 dimension flatten vectors at the
moment, which can be improved in the future when it’s needed.