Building conda packages#


This tutorial describes how to use conda-build to create conda packages on Windows, macOS, and Linux using the examples of SEP and GDAL. Additional Windows-specific instructions are provided in the Toolkit section.

The final built packages from this tutorial are available on (formerly known as Anaconda Cloud):

This tutorial also describes writing recipes. You can see the final SEP recipe and the GDAL recipe on GitHub in the conda-build documentation repository.

Who is this for?#

This tutorial is for Windows, macOS, and Linux users who wish to build more complex conda packages. This tutorial will involve building scientific packages, which require compilers for several different Python versions.

Before you start#

Before you start, make sure you have installed:


Microsoft Visual Studio#

In the standard practices of the conda developers, conda packages for different versions of Python are each built with their own version of Visual Studio (VS):

  • Python 2.7 packages with Visual Studio 2008

  • Python 3.4 packages with VS 2010

  • Python 3.5 packages with VS 2015, (default) 2017

  • Python 3.6 packages with VS 2015, (default) 2017

  • Python 3.7 packages with VS 2015, (default) 2017

Using these versions of VS to build packages for each of these versions of Python is also the practice used for the official builds of Python. Currently VS 2008 and VS 2010 are available only through resellers, while VS 2015 and VS 2017 can be purchased online from Microsoft. Note there is also a community edition of VS 2015 and VS 2017 which may be used.

Alternatives to Microsoft Visual Studio#

There are free alternatives available for each version of the VS compilers:

The MS Visual C++ Compiler for Python 2.7 and the Microsoft Windows SDK for Windows 7 and .NET Framework 4 are both reasonably well tested. Conda-build is carefully tested to support these configurations, but there are known issues with the CMake build tool and these free VS 2008 and 2010 alternatives. In these cases, you should prefer the "NMake Makefile" generator, rather than a Visual Studio solution generator.

Windows versions#

You can use any recent version of Windows. These examples were built on Windows 10.

Other tools#

Some environments initially lack tools such as patch or Git that may be needed for some build workflows.

On Windows, these can be installed with conda using the following command:

$ conda install git m2-patch

On macOS and Linux, replace m2-patch with patch.

Developing a build strategy#

Conda recipes are typically built with a trial-and-error method. Sometimes, the first attempt to build a package will fail with compiler or linker errors, often caused by missing dependencies. The person writing the recipe then examines these errors and modifies the recipe to include the missing dependencies, usually as part of the meta.yaml file. Then the recipe writer attempts the build again and, after a few of these cycles of trial and error, the package builds successfully.

Building with a Python version different from your Miniconda installation#

Miniconda2 and Miniconda3 can each build packages for either Python 2 or Python 3 simply by specifying the version you want. Miniconda2 includes only Python 2 and Miniconda3 includes only Python 3.

Installing only one makes it easier to keep track of the builds, but it is possible to have both installed on the same system at the same time. If you have both installed, use the where command on Windows, or the which command on macOS or Linux, to see which version comes first on PATH, since this is the one you will be using:

$ where python

To build a package for a Python version other than the one in your Miniconda installation, use the --python option in the conda-build command. For example, to build a Python 3.5 package with Miniconda2:

$ conda-build recipeDirectory --python=3.5


Replace recipeDirectory with the name and path of your recipe directory.

Automated testing#

After the build, if the recipe directory contains a test file. This test file is named run_test.bat on Windows, on macOS or Linux, or on any platform. The file runs to test the package and any errors that surface are reported. After seeing "check the output," you can also test if this package was built by using the command:

$ conda build --test <path to package>.tar.bz2


Use the "test" section of the meta.yaml file to move data files from the recipe directory to the test directory when the test is run.

Building a SEP package with conda and Python 2 or 3#

The SEP documentation states that SEP runs on Python 2 and 3, and it depends only on NumPy. Searching for SEP on PyPI shows that there is already a PyPI package for SEP.

Because a PyPI package for SEP already exists, the conda skeleton command can make a skeleton or outline of a conda recipe based on the PyPI package. The recipe outline can then be completed manually and conda can build a conda package from the completed recipe.

Install Visual Studio#

If you have not already done so, install the appropriate version of Visual Studio:

  • For Python 3 — Visual Studio 2017:

    1. Choose the "Custom install" option.

    2. Under Programming Languages, choose to install Visual C++.

  • For Python 2 — Visual Studio 2008:

    1. Choose the "Custom install" option.

    2. Choose to install X64 Compilers and Tools. Install Service Pack 1.

Make a conda skeleton recipe#

  1. Run the skeleton command:

    $ conda skeleton pypi sep

    The skeleton command installs into a newly-created directory called sep.

  2. Go to the sep directory to view the files:

    $ cd sep

    You will see that one skeleton file has been created: meta.yaml

Edit the skeleton files#

For this package, bld.bat and need no changes. You need to edit the meta.yaml file to add the dependency on NumPy and add an optional test for the built package by importing it. For more information about what can be specified in meta.yaml, see Defining metadata (meta.yaml).

  1. In the requirements section of the meta.yaml file, add a line that adds NumPy as a requirement to build the package.

  2. Add a second line to list NumPy as a requirement to run the package.

  3. Set the NumPy version to the letters x.x.

  4. Make sure the new line is aligned with - python on the line above it, so as to ensure proper yaml format.


    - python
    - numpy     x.x

    - python
    - numpy     x.x

Notice that there are two types of requirements, host and run (build is another valid parameter, but is not shown in this example). host represents packages that need to be specific to the target platform when the target platform is not necessarily the same as the native build platform. run represents the dependencies that should be installed when the package is installed.


Using the letters x.x instead of a specific version such as 1.11 pins NumPy dynamically, so that the actual version of NumPy is taken from the build command. Currently, NumPy is the only package that can be pinned dynamically. Pinning is important for SEP because this package uses NumPy's C API through Cython. That API changes between NumPy versions, so it is important to use the same NumPy version at runtime that was used at build time.

Optional: Add a test for the built package#

Adding this optional test will test the package at the end of the build by making sure that the Python statement import sep runs successfully:

  1. Add - sep, checking to be sure that the indentation is consistent with the rest of the file:

      # Python imports
        - sep

Build the package#

Build the package using the recipe you just created:

$ conda build sep

Check the output#

  1. Check the output to make sure that the build completed successfully. The output contains the location of the final package file and a command to upload the package to Anaconda Cloud. The output will look something like:

    # Automatic uploading is disabled
    # If you want to upload package(s) to later, type:
    anaconda upload /Users/builder/miniconda3/conda-bld/osx-64/sep-1.0.3-np111py36_0.tar.bz2
    # To have conda build upload to automatically, use
    # $ conda config --set anaconda_upload yes
    anaconda_upload is not set.  Not uploading wheels: []
    Resource usage summary:
    Total time: 0:00:56.4
    CPU usage: sys=0:00:00.7, user=0:00:07.0
    Maximum memory usage observed: 220.1M
    Total disk usage observed (not including envs): 3.9K
    Source and build intermediates have been left in /Users/builder/miniconda3/conda-bld.
    There are currently 437 accumulated.
    To remove them, you can run the ```conda build purge``` command
  1. If there are any linker or compiler errors, modify the recipe and build again.

Building a GDAL package with conda and Python 2 or 3#

To begin, install Anaconda or Miniconda and conda-build. If you are using a Windows machine, also use conda to install Git and m2-patch.

$ conda install git
$ conda install m2-patch

Because GDAL includes C and C++, building it on Windows requires Visual Studio. This procedure describes how to build a package with Python 2 or Python 3. Follow the instructions for the version with which you want to build.

To build a GDAL package:

  1. Install Visual Studio:

    • Choose "Custom install".

    • Under "Programming Languages", select workloads that come from Visual Studio so that you can choose the Desktop Development with C++ and Universal Platform C.

    • Choose "Custom install".

    • Choose to install X64 Compilers and Tools.

    • Install Visual Studio 2008 Service Pack 1.

  2. Install Git. Because the GDAL package sources are retrieved from GitHub for the build, you must install Git:

    $ conda install git m2-patch conda-build
  3. Get gdal-feedstock. For the purpose of this tutorial, we will be using a recipe from Anaconda:

    $ git clone
  4. Use conda-build to build the gdal-feedstock:

    $ conda build gdal-feedstock
  5. Check the output to make sure that the build completed successfully. The output also contains the location of the final package file and a command to upload the package to Cloud. For this package in particular, there should be two packages outputted: libgdal and GDAL.

  6. In case of any linker or compiler errors, modify the recipe and run it again.

Let's take a closer look at what's happening inside the gdal-feedstock, specifically in the meta.yaml file.

The first interesting bit happens under source in the patches section:

  - 0001-windowshdf5.patch
  # Use multiple cores on Windows.
  - 0002-multiprocessor.patch
  # disable 12 bit jpeg on Windows as we aren't using internal jpeg
  - 0003-disable_jpeg12.patch

This section is basically saying "when this package is being built on a Windows platform, apply the following patch files". Notice that the patch files are in the patches directory of the recipe. These patches will only be applied to Windows since the # [win] selector is applied to each of the patch entries. For more about selectors, see Preprocessing selectors.

In the requirements section, notice how there are both a build and host set of requirements (run is another valid parameter, but is not shown in this example). For this recipe, all the compilers required to build the package are listed in the build requirements. Normally, this section will list out packages required to build the package. GDAL requires CMake on Windows, as well as C compilers. Notice that the C compilers are pulled into the recipe using the syntax {{ compiler('c') }}. Since conda-build 3, conda-build defines a jinja2 function compiler() to specify compiler packages dynamically. So, using the compiler('c') function in a conda recipe will pull in the correct compiler for any build platform. For more information about compilers with conda-build see compiler-tools.

Also note that the compilers used by conda-build can be specified using a conda_build_config.yaml. For more information about how to do that, see Using your customized compiler package with conda-build 3.

Notice that this package has an outputs section. This section is a list of packages to output as a result of building this package. In this case, the packages libgdal and GDAL will be built. Similar to a normal recipe, the outputs can have build scripts, tests scripts, and requirements specified. For more information on how outputs work, see the Outputs section.

Now, let's try to build GDAL against some build matrix. We will specify building against Python 3.7 and 3.5 using a conda-build config. Add the following to your conda_build_config.yaml:

   - 3.7
   - 3.5

Now you can build GDAL using conda-build with the command:

$ conda build gdal-feedstock

Or explicitly set the location of the conda-build variant matrix:

$ conda build gdal-feedstock --variant-config-file conda_build_config.yaml

If you want to know more about build variants and conda_build_config.yaml, including how to specify a config file and what can go into it, take a look at Creating conda-build variant config files.