Skip to content

Latest commit

 

History

History
363 lines (275 loc) · 15.3 KB

GUIDELINES.md

File metadata and controls

363 lines (275 loc) · 15.3 KB

Guidelines for bioconda recipes

Requirements of a good bioconda recipe

  • Follows the Filesystem Hierarchy Standard (FSH). This is most important for libraries and Java packages; for these cases use one of the recipes below as a guideline.

  • Has adequate tests (see "Testing" section below)

  • License allows redistribution and license is indicated in meta.yaml

  • Does not already exist in the main Anaconda channel or the r channel. Exceptions: Anaconda version is too old or some sort of patch is required that is not in the main Anaconda channel (TODO: examples of the latter?)

  • If the recipe installs custom wrapper scripts, usage notes should be added to extra -> notes in the meta.yaml.

  • If uploading of an unreleased version is necessary, please follow the versioning scheme of conda for pre- and post-releases (e.g. using a, b, rc, and dev suffixes, see here).

Examples

The following recipes serve as examples of good recipes that can be used as guides or templates when developing new recipes.

Python

Use conda skeleton pypi <packagename> where packagename is a package available on PyPI. Either run that command in the recipes dir or move the recipe it creates to recipes. Typically can be used without modification, though dependencies may also need recipes.

If the recipe was created using conda skeleton pypi, then the default test is probably sufficient. The exception is when the package also installs a command-line tool, in which case that should be tested as well.

  • typical import check: pysam
  • import and command-line tests: chanjo

R (CRAN)

Use conda skeleton cran <packagename> where packagename is a package available on CRAN and is case-sensitive. Either run that command in the recipes dir or move the recipe it creates to recipes. The recipe name will have an r- prefix and will be converted to lowercase. Typically can be used without modification, though dependencies may also need recipes.

If the recipe was created using conda skeleton cran or the scripts/bioconductor_skeleton.py script, the default test is probably sufficient.

  • typical R recipe from CRAN: r-locfit
  • recipe for R package not on CRAN, also with patch: spp

R (Bioconductor)

Use scripts/bioconductor/bioconductor_skeleton.py <packagename> where packagename is a case-sensitive package available on Bioconductor. The recipe name will have a bioconductor- prefix and will be converted to lowercase. Typically can be used without modification, though dependencies may also need recipes. Recipes for dependencies with an r- prefix should be created using conda skeleton cran; see above.

Java

Add a wrapper script if the software is typically called via java -jar .... For example, fastqc already had a wrapper script, but peptide-shaker did not.

JAR files should go in $PREFIX/share/$PKG_NAME-$PKG_VERSION-$PKG_BUILDNUM. A wrapper script should be placed here as well, and symlinked to $PREFIX/bin.

Perl

Use conda skeleton cpan <packagename> to build a recipe for Perl and place the recipe in the recipes dir. The recipe will have the perl- prefix.

The recipe as generated by conda skeleton cpan must be changed. The run and build requirements must specify perl-threaded instead of perl. Since some bioconda packages depend on a version of Perl compiled with threading support, a choice was made to have all recipes use perl-threaded to avoid maintaining multiple versions of each Perl module.

An example of such a package is perl-module-build.

Alternatively, you can additionally ensure the build requirements for the recipe include perl-app-cpanminus, and then the build.sh script can be simplified. An example of this simplification is perl-time-hires.

If the recipe was created with conda skeleton cpan, the tests are likely sufficient. Otherwise, test the import of modules (see the imports section of the meta.yaml files in above examples).

C/C++

In general, standard make should work. Other build tools (e.g., autoconf) and compilers (e.g., gcc) should be specified in the build requirements.

We have not yet decided whether to have gcc as a conda package or to assume it is in the build environment. Until this decision is made, please add gcc (for Linux packages) and llvm (for OSX packages) to the meta.yaml as follows:

requirements:
  build:
    - gcc   # [not osx]
    - llvm  # [osx]

  run:
    - libgcc    # [not osx]

If the package uses zlib, then please see the ZLIB section below.

If your package links dynamically against a particular library, it is often necessary to pin the version against which it was compiled, in order to avoid ABI incompatibilities. Instead of hardcoding a particular version in the recipe, we use jinja templates to achieve this. For example, bioconda provides an environnmnet variable CONDA_BOOST that contains the current major version of boost. You should pin your boost dependency against that version. An example is the salmon recipe. You find the libraries you can currently pin in scripts/env_matrix.yml. If you need to pin another library, please notify @bioconda/core, and we will set up a corresponding environment variable.

It's not uncommon to have difficulty compiling package into a portable conda package. Since there is no single solution, here are some examples of how bioconda contributors have solved compiling issues to give you some ideas on what to try:

  • ococo edits the source in build.sh to accommodate the C++ compiler on OSX
  • muscle patches the makefile on OSX so it doesn't use static libs
  • metavelvet, eautils, preseq have several patches to their makefiles to fix LIBS and INCLUDES, INCLUDEARGS, and CFLAGS
  • mapsplice includes an older version of samtools; the included samtools' makefile is patched to work in conda envs.
  • mosaik has platform-specific patches -- one removes -static on linux, and the other sets BLD_PLATFORM correctly on OSX
  • mothur and soapdenovo have many fixes to makefiles

General command-line tools

If a command-line tool is installed, it should be tested. If it has a shebang line, it should be patched to use /usr/bin/env for more general use.

An example of this is fastq-screen.

For command-line tools, running the program with no arguments, checking the programs version (e.g. with -v) or checking the command-line help is sufficient if doing so returns an exit code 0. Often the output is piped to /dev/null to avoid output during recipe builds.

Examples:

If a package depends on Python and has a custom build string, then py{{CONDA_PY}} must be contained in that build string. Otherwise Python will be automatically pinned to one minor version, resulting in dependency conflicts with other packages. See mapsplice for an example of this.

Metapackages

Metapackages tie together other packages. All they do is define dependencies. For example, the hubward-all metapackage specifies the various other conda packages needed to get full hubward installation running just by installing one package. Other metapackages might tie together conda packages with a theme. For example, all UCSC utilities related to bigBed files, or a set of packages useful for variant calling.

For packages that are not anchored to a particular package (as in the last example above), we recommended semantic versioning starting at 1.0.0 for metapackages.

Other examples

Examples of somewhat non-standard recipes, in no particular order:

  • blast has an OS-specific installation -- OSX copies binaries while on Linux it is compiled.
  • graphviz has an OS-specific option to configure
  • crossmap removes libs that are shipped with the source distribution
  • hisat2 runs 2to3 to make it Python 3 compatible, and copies over individual scripts to the bin dir
  • krona has a post-link.sh script that gets called after installation to alert the user a manual step is required
  • htslib has a small test script that creates example data and runs multiple programs on it
  • spectacle runs 2to3 to make the wrapper script Python 3 compatible, patches the wrapper script to have a shebang line, deletes example data to avoid taking up space in the bioconda channel, and includes a script for downloading the example data separately.
  • gatk is a package for licensed software that cannot be redistributed. The package installs a placeholder script (in this case doubling as the jar wrapper) to alert the user if the program is not installed, along with a separate script (gatk-register) to copy in a user-supplied archive/binary to the conda environment

Name collisions

In some cases, there may be a name collision when writing a recipe. For example the wget recipe is for the standard command-line tool. There is also a Python package called wget on PyPI. In this case, we prefixed the Python package with python- (see python-wget). A similar collision was resolved with weblogo and python-weblogo.

If in doubt about how to handle a naming collision, please submit an issue.

Fixes for missing URLs

Sometimes, URLs to source code may disappear -- particularly with R and Bioconductor packages, where only the very latest source code is hosted on the Bioconductor servers. This means that rebuilding an older package version fails because the source can't be found.

The solution to this is the Cargo Port, developed and maintained by the Galaxy team. The Galaxy Jenkins server performs daily archives of the source code of packages in bioconda, and makes these tarballs permanently available in Cargo Port. If you try rebuilding a recipe and the source seems to have disappeared, do the following:

  • search for the package and version at https://depot.galaxyproject.org/software/
  • add the URL listed in the "Package Version" column to your meta.yaml file as another entry in the source: url section.
  • add the corresponding sha256 checksum displayed upon clicking the Info icon in the "Help" column to the source: section.

For example, if this stopped working:

source:
  fn: argh-0.26.1.tar.gz
  url: https://pypi.python.org/packages/source/a/argh/argh-0.26.1.tar.gz
  md5: 5a97ce2ae74bbe3b63194906213f1184

then change it to this:

source:
  fn: argh-0.26.1.tar.gz
  url:
    - https://pypi.python.org/packages/source/a/argh/argh-0.26.1.tar.gz
    - https://depot.galaxyproject.org/software/argh/argh_0.26.1_src_all.tar.gz
  md5: 5a97ce2ae74bbe3b63194906213f1184
  sha256: 06a7442cb9130fb8806fe336000fcf20edf1f2f8ad205e7b62cec118505510db

Tests

An adequate test must be included in the recipe. An "adequate" test depends on the recipe, but must be able to detect a successful installation. While many packages may ship their own test suite (unit tests or otherwise), including these in the recipe is not recommended since it may timeout the build system on Travis-CI.

Note that a test must return an exit code of 0. The test can be in the test field of meta.yaml, or can be a separate script (see the relevant conda docs for testing).

It is recommended to pipe unneeded stdout/stderr to /dev/null to avoid cluttering the output in the Travis-CI build environment.

Link and unlink scripts (pre- and post- install hooks)

It is possible to include scripts that are executed before or after installing a package, or before uninstalling a package. These scripts can be helpful for alerting the user that manual actions are required after adding or removing a package. For example, a post-link.sh script may be used to alert the user that he or she will need to create a database or modify a settings file. Any package that requires a manual preparatory step before it can be used should consider alerting the user via an echo statement in a post-link.sh script. These scripts may be added at the same level as meta.yaml and build.sh:

  • pre-link.sh is executed prior to linking (installation). An error causes conda to stop.
  • post-link.sh is executed after linking (installation). When the post-link step fails, no package metadata is written, and the package is not considered installed.
  • pre-unlink.sh is executed prior to unlinking (uninstallation). Errors are ignored. Used for cleanup.

These scripts have access to the following environment variables:

  • $PREFIX The install prefix
  • $PKG_NAME The name of the package
  • $PKG_VERSION The version of the package
  • $PKG_BUILDNUM The build number of the package

Versions

If an older version is required, put it in a subdirectory of the recipe. Examples of this can be found in bowtie2, bx-python, and others.

ZLIB

If a package depends on zlib, then it will most likely also depend on gcc (on Linux) and llvm (on OSX). The meta.yaml requirements section should therefore at least have the following for a recipe that supports both Linux and OSX:

requirements:
  build:
    - gcc   # [not osx]
    - llvm  # [osx]
    - zlib

  run:
    - libgcc    # [not osx]
    - zlib

When building the package, you may get an error saying that zlib.h can't be found -- despite having zlib listed in the dependencies. The reason is that the location of zlib often has to be specified in the build.sh script, which can be done like this:

export CFLAGS="-I$PREFIX/include"
export LDFLAGS="-L$PREFIX/lib"

Or sometimes:

export CPATH=${PREFIX}/include

Sometimes Makefiles may specify these locations, in which case they need to be edited. See the samtools recipe for an example of this. It may take some tinkering to get the recipe to build; if it doesn't seem to work then please submit an issue.