# PIG 13 - Gammapy dependencies and distribution¶

• Author: Christoph Deil, Axel Donath, Régis Terrier, Brigitta Sipocz
• Created: June 6, 2019
• Accepted: Sep 9, 2019
• Status: accepted
• Discussion: GH 2218

## Abstract¶

Now that we have a good way to distribute Gammapy and to ship a science tool environment to users, we propose to drop support and testing for old versions of dependencies, and alternative distribution channels. Concretely, we propose to require Python 3.6, Numpy 1.16 and Astropy 3.2 starting with Gammapy v0.14, and to remove the Macports installation instructions from the Gammapy documentation.

We think the impact for users is small (none for most), but the benefit for Gammapy developers and maintainers is big, allowing us to progress more quickly.

If you use Gammapy and need to run on old machines or exotic platforms, and this change doesn’t work for you, let us know!

## Introduction¶

Since Gammapy v0.7, released in Feb 2018, the recommended way to install Gammapy and its dependencies has been via conda. This has worked well, allowing us to ship a reproducible science tool environment with up-to-date versions to all Gammapy users for each Gammapy release.

Each stable release for Gammapy is first published as a source distribution on https://pypi.org/project/gammapy/. Then conda binaries for Linux, macOS and Windows will be built via conda-forge, and uploaded to https://anaconda.org/conda-forge/gammapy. Finally we write a conda environment specification file that we publish on gammapy.org, and the end-user installation instructions look like this:

curl -O https://gammapy.org/download/install/gammapy-0.13-environment.yml
conda env create -f gammapy-0.13-environment.yml
conda activate gammapy-0.13


### Dependencies¶

For the rest of this document, we would like to define and describe what we mean when talking about “required” and “optional” dependencies for Gammapy:

• the required dependencies are the Python packages that get automatically installed when running pip install gammapy or conda install gammapy -c conda-forge.
• the optional dependencies are the ones listed in the conda environment specification, that are not already required dependencies.

For example Numpy and Astropy are required dependencies of Gammapy, and matplotlib, Jupyter or Naima are optional dependencies. Complete tables of all required and optional dependencies that we have at the moment are given below.

The choice which dependencies are required or optional (or neither) is something we make in the metadata of the Gammapy source distribution. Our reasoning is to declare a dependency as required if it’s needed by the vast majority of Gammapy users, and for optional dependencies, we include all packages that are used somewhere within the Gammapy package, or in examples and tutorials in the Gammapy documentation.

We note that these two sets of dependencies are just the default, recommended sets, it is possible to install Gammapy even without installing the required dependencies, and of course users can install and use Gammapy together with other Python packages, e.g. scikit-learn or whatever they like.

### Distributions¶

There are many distribution channels for Gammapy just as is the case for most open-source software. E.g. at this time there is a Debian and Macports package for Gammapy. conda is not the only, but it is the only fully supported distribution channel for Gammapy. The reason for this is that currently manpower and expertise in the Gammapy team is limited, and conda provides, as far as we know, a solution that works for all users. At this time, no-one from the Gammapy is using Macports any more, for pip there are no binary wheels published yet for Gammapy and astropy-regions, Debian only works for Debian users and has a longer update cycle compared to the current Gammapy 2 month release cycle, and e.g. for Homebrew no-one packaged Gammapy so far.

## Required dependencies¶

We propose to update the Gammapy required dependencies as shown in the following table (the release dates for the packages are shown in parentheses are were obtained from https://pypi.org/).

Dependency Gammapy 0.13 Gammapy 0.14
Python 3.5 (Sep 2015) 3.6 (Dec 2016)
Numpy 1.10 (May 2016) 1.16 (Jan 2019)
Scipy 0.15 (Jan 2015) 1.2 (Dec 2018)
Astropy 2.0 (Jul 2017) 3.2 (Jun 2019)
regions 0.4 (Jun 2019) 0.5 (Sep 2019)
pyyaml unclear 5.1 (Mar 2019)
click unclear 7.0 (Sep 2018)
jsonschema 3.0 (Feb 2019)

We already mentioned the possibility to drop Python 3.5 support in PIG 3 - Plan for dropping Python 2.7 support. One reason is that Anaconda and conda-forge (our main distribution channel, used in our testing continuous integration setup) only contains Python 2.7, 3.6 and 3.7 at this point (and 3.8 added in fall 2018), i.e. testing on Python 3.5 is already extra effort. Also, Python 3.6 contains some nice new features that developers can use. E.g. Sunpy or ctapipe already require Python 3.6 or later.

A major motivation to update to very recent versions is that the regions package is still under development (see PIG 10 - Regions). In Gammapy 0.13 we require regions 0.4, and we plan to make a regions 0.5 release with further features and fixes in September, and to require that for Gammapy 0.14.

For pyyaml, we would like to use the recent version, since it allows writing with preserved dictionary key order, giving a nice output, whereas previously it always sorted keys alphabetically on YAML write.

The current Gammapy command line interface is using click. Whether to keep this or whether to use something else will be discussed in PIG 12. For now, we propose to keep things as-is, and only specify a minimum version that we’ll test, although in practice we didn’t have any version-dependent issues with click in the past years.

We plan to use jsonschema to validate YAML config files. It’s a small, pure-Python dependency.

## Optional dependencies¶

We propose to update the Gammapy optional dependencies as shown in the following table (the release dates for the packages are shown in parentheses are were obtained from https://pypi.org/).

Dependency Gammapy 0.13 Gammapy 0.14
ipython 7.3 (Feb 2019) 7.6 (Jun 2019)
jupyter 1.0 (n/a) 1.0 (n/a)
jupyterlab 0.35 (Oct 2018) 1.0 (Jun 2019)
matplotlib 2.1 (Oct 2017) 3.0 (Sep 2018)
pandas 0.24 (Jan 2019) 0.25 (Jul 2019)
healpy 1.11 (Aug 2017) 1.12 (Jun 2018)
reproject 0.4 (Jan 2018) 0.5 (Jun 2019)
uncertainties 3.0 (Aug 2016) 3.1 (May 2019)
iminuit 1.3 (Jul 2018) 1.3.7 (Jun 2019)
sherpa 4.11 (Feb 2019) 4.11 (Feb 2019)
naima 0.8 (Dec 2016) 0.8.3 (Nov 2018)
emcee 2.2 (Jul 2016)
corner 2.0 (May 2016)
parfive 1.0 (May 2019)

We plan to use parfive for tutorial notebook and example dataset file download (features parallel download and a progress bar).

## Distributions¶

The situation concerning distribution was described in the introduction above, and some alternatives and future work are mentioned in the outlook and alternatives sections below.

We propose to remove the Macports installation page from the Gammmapy documentation, and to just leave a mention that it’s outdated and unsupported on the page on “other ways to install Gammapy”. Short sections or pages about pip and Debian will remain at the end of the installation instructions, for advanced users.

For conda, we will improve the installation instructions, explaining the difference between using the recommended environment and a plain conda install -c conda-forge gammapy, and add lists of required and optional dependencies as above to the Gammapy installation documentation.

## Outlook¶

This PIG describes only the status and very short term plan for Gammapy v0.14 and v1.0! We expect that dependencies and distribution will evolve in the coming years.

E.g. we might add Numba as a dependency, or we might start to fully support pip, if binary wheels for Gammapy and all dependencies are available (there is a recent effort to implement this).

Somewhat related to this PIG: we plan to modernise Gammapy packaging following the recommendations in APE 17. Work on this has already started in GH 2279.

We could support other ways to ship Gammapy, e.g. Homebrew (works on MacOS and also Linux now), or Docker images (works anywhere, self-contained).

## Alternatives¶

Since we have a working setup already, we could do nothing, which seems nice at first, but really means continued time sink and suffering by the Gammapy maintainers of the continuous integration testing system and the release manager.

For conda, we should strongly consider introducing a gammapy-all metapackage instead of the current gammapy-environment.yml shipped via gammapy.org. That’s what e.g. the Fermi tools and glueviz and others do, and they like that solution. It does have the advantage ob using a conda-native solution to the question how to ship an environment. This could be prototyped any time, in parallel with the existing solution, to gain familiarity with metapackages.

• Remove extras_require in setup.py (don’t attempt to maintain the list of optional dependencies there)