FAQs¶
How do I speed up package installation?¶
Speedup option 1: use mamba
¶
mamba is a drop-in replacement for conda that uses a faster dependency solving library and parts reimplemented in C++ for speed. Install it just into the base environment so that it’s always available, like this:
conda install mamba -n base -c conda-forge
Then use mamba
instead of conda
.
For example, instead of conda install
, use mamba install
. Instead of
conda env create
use mamba env create
, and so on. mamba
also uses
the same configuration as conda
, so you don’t need to reconfigure the
channels.
Note
Installing mamba
into the base environment (-n base
in the command
above) means that it does not need to be installed into each subsequent
environment you create.
Speedup option 2: use environments strategically¶
Here are several ways you can use environments to minimize the time spent on solving dependencies, which typically is what takes the longest amount of time:
Keep the base environment small.
If you install everything into the same environment (e.g. the
base
environment, which is used any time you don’t otherwise specify an environment), then whenever you add or update packages to it, the solver has to do a lot of work to make sure all of the many packages are mutually compatible with each other.Use smaller environments.
Fewer packages means less work for the solver. Try to use environments only containing what you need for a particular project or task.
Pin dependencies.
Sometimes pinning dependencies to a specific version can speed up the solving, since it reduces the search space for the solver. In some cases this may backfire though. For example, you can’t pin an older version of R and also use newer R packages that don’t support that version of R.
Create an environment from a file with all dependencies.
Creating an environment with all dependencies at once can be faster than incrementally adding packages to an existing environment. For example
conda create -n myenv --file requirements.txt
, orconda env create --file env.yaml
.Use strict channel priority.
Ensure that you’ve run
conda config --set channel_priority strict
to respect the configured channel order, as recommended in the setup instructions. This can also speed up the solving.
What versions are supported?¶
Operating Systems: Bioconda supports Linux (x86_64 and aarch64/arm64) and macOS (x86_64 and arm64). Windows is not supported.
Python: Bioconda currently supports Python 3.8, 3.9, 3.10, 3.11, and 3.12 (see “pinned packages” below for where this is configured). There are still packages in the Bioconda channel for earlier versions of Python (2.7, 3.6, and 3.7), but new packages are not built for these versions.
Packages which declare noarch: python
and only depend on packages that also
declare noarch: python
can be installed in an environment with any version of
Python they say they can support. However many Python packages in Bioconda
depend on other Bioconda packages with architecture specific builds, such as
pysam
, and so do not meet this criteria.
Changed on 2022-09-01: Python 3.10 support started in Aug 2022
Changed on 2023-05-01: Python 2.7, 3.6, 3.7 support were dropped for new recipes in May 2023
Changed on 2024-06-15: Python 3.11, 3.12 support started in June 2024
Globally-pinned versions: See How does global pinning work? for details.
Unsupported versions: If there is a version of a dependency you wish to build against that Bioconda does not currently support, please reach out to the Bioconda Gitter for more information about if supporting that version is feasible, if work on that is already being done, and how you can help.
To find out against which version you can pin a package, e.g. x.y.* or x.* please use ABI-Laboratory.
How do I keep track of environments?¶
You can view your created environments with conda env list
.
Note that if keeping track of different environment names
becomes a burden, you can create an environment in the same directory as
a project with the -p
argument, e.g.,
conda create -p ./env --file requirements.txt
and then activate the environment with
conda activate ./env
This also works quite well in a shared directory so everyone can use (and maintain) the same environment.
What’s the difference between Anaconda, conda, Miniconda, mamba, Mambaforge, micromamba?¶
This blog post from Anaconda gives a lot of context on the Anaconda/conda ecosystem.
conda is the name of the package manager, which is what runs when you call, e.g.,
conda install
.mamba is a drop-in replacement for conda.
Anaconda is a large installation including Python, conda, and a large number of packages.
Miniconda just has conda and its dependencies (in contrast to the larger Anaconda distribution).
Miniforge is like miniconda, but with the conda-forge channel preconfigured and all packages coming from the conda-forge and not the
defaults
channel. It also now has mamba and libmamba included.Mambaforge is like miniforge, but has mamba installed into the base environment. While not strictly deprecated, its use is discouraged as of Sept 2023 (see miniforge README).
Micromamba is not a conda distribution. Rather, it is a minimal binary that has roughly the same commands as mamba, so that a single executable (rather than an entire Python installation required for conda itself) can be used to create environments. Micromamba is currently still experimental.
The Anaconda Python distribution started out as a bundle of scientific Python packages that were otherwise difficult to install. It was created by ContinuumIO and remains the easiest way to install the full scientific Python stack.
Many packaging problems had to be solved in order to provide all of that software in Anaconda in a cross-platform bundle, and one of the tools that came out of that work was the conda package manager. So conda is part of th Anaconda Python distribution. But conda ended up being very useful on its own and for things other than Python, so ContinuumIO spun it out into its own separate open-source package.
Conda became very useful for setting up lightweight environments for testing code or running individual steps of a workflow. To avoid needing to install the entire Anaconda distribution each time, the Miniconda installer was created. This installs only what you need to run conda itself, which can then be used to create other environments. So the “mini” in Miniconda means that it’s a fraction of the size of the full Anaconda installation.
Then the conda-forge channel gained popularity. Miniforge was developed to quickly and easily get a conda-forge-ready conda installation. Then as mamba gained popularity, the Mambaforge variant was created.
Even with those easier methods, sometimes the entire base Python installation that comes with conda/mamba was too much overhead. Micromamba has a single binary that is very fast to install, and is perfect for CI environments.
So: conda is a package manager, Anaconda is a scientific Python distribution that also includes conda, and the rest are other flavors of getting a conda/mamba installation.
What’s the difference between a recipe and a package?¶
A recipe is a directory containing small set of files that defines name,
version, dependencies, and URL for source code. A recipe typically contains
a meta.yaml
file that defines these settings and a build.sh
script that
builds the software.
A recipe is converted into a package by running conda-build
on the recipe.
A package is a bgzipped tar file (.tar.bz2
) that contains the built
software in expected subdirectories, along with a list of what other packages
are dependencies. For example, a conda package built for a Python package would
end up with .py
files in the lib/python3.8/site-packages/
directory inside the tarball, and would specify (at least) Python as
a dependency.
Packages are uploaded to anaconda.org so that users can install them
with conda install
.
See also
The Conda package specification
has details on exactly
what a package contains and how it is installed into an
environment.
Why are Bioconductor data packages failing to install?¶
When creating an environment containing Bioconductor data packages, you may get errors like this:
ValueError: unsupported format character 'T' (0x54) at index 648
The actual error will be somewhere above that, with something like this (here,
it’s for the bioconductor-org.hs.eg.db=3.14.0=r41hdfd78af_0
package):
message:
post-link script failed for package bioconda::bioconductor-org.hs.eg.db-3.14.0-r41hdfd78af_0
location of failed script: /Users/dalerr/env/bin/.bioconductor-org.hs.eg.db-post-link.sh
==> script messages <==
<None>
==> script output <==
stdout: ERROR: post-link.sh was unable to download any of the following URLs with the md5sum ef7fc0096ec579f564a33f0f4869324a:
https://bioconductor.org/packages/3.14/data/annotation/src/contrib/org.Hs.eg.db_3.14.0.tar.gz
https://bioarchive.galaxyproject.org/org.Hs.eg.db_3.14.0.tar.gz
https://depot.galaxyproject.org/software/bioconductor-org.hs.eg.db/bioconductor-org.hs.eg.db_3.14.0_src_all.tar.gz
To fix it, you need to adjust the requirements. If you had this as a requirement:
bioconductor-org.hs.eg.db=3.14.0=r41hdfd78af_0
then increase the build number on the end, here from _0
to _1
:
bioconductor-org.hs.eg.db=3.14.0=r41hdfd78af_1
or, relax the exact build constraint while keeping the package version the same:
bioconductor-org.hs.eg.db=3.14.0
and then re-build your environment.
The reason this is happening is a combination of factors. Early on in Bioconda’s history we made the decision that pure data packages – like Bioconductor data packages, which can be multiple GB in size – would not be directly converted into conda packages. That way, we could avoid additional storage load on Anaconda’s servers since the data were already available from Bioconductor, and we could provide a mechanism to use the data packages within an R environment living in a conda environment. This mechanism is a post-link.sh script for the recipe.
When a user installs the package via conda, the GB of data aren’t in the
package. Rather, the URL pointing to the tarball is in the post-link script,
and the script uses curl
to download the package from Bioconductor and
install into the conda environment’s R library. We also set up separate
infrastructure to archive data packages to other servers, and these archive
URLs were also stored in the post-link scripts as backups.
The problem is that back then, we assumed that URLs would be stable and we did
not use the -L
argument for curl in post-link scripts.
Recently Bioconductor packages have moved to a different server (XSEDE/ACCESS).
The old URL, the one hard-coded in the post-link scripts, is correctly now
a redirect to the new location. But without -L
, the existing recipes and
their post-link scripts cannot follow the redirect! Compounding this, the
archive URLs stopped being generated, so the backup strategy also failed.
The fix was to re-build all Bioconductor data packages and include the -L
argument, allowing them to follow the redirect and correctly install the
package. Conda packages have the idea of a “build number”, which allows us to
still provide the same version of the package (3.14.0 in the example above) but
packaged differently (in this case, with a post-link script that works in
Bioconductor’s current server environment).
Reproducibility is hard. We are trying our best, and conda is an amazing resource. But the fact that a single entity does not (and should not!) control all code, data, packages, distribution mechanisms, and installation mechanisms, means that we will always be at risk of similar situations in the future. Hopefully we are guarding better against this particular issue, but see Grüning et al 2018 (especially Fig 1) for advice on more reproducible strategies you can use for your own work.
What’s the difference between a build number and a package version?¶
A package version is the version of the tool. A tool can possibly be packaged multiple times, even though the underlying tool doesn’t change. In such a case, the package version remains unchanged, but the build number chances.
The Bioconductor data packages described above are one example of what would cause a change in build number (i.e., adding a single argument to a post-installation script). Other times, a package might have omitted an executable that should have been included, so a new build for the same version is created that fixes that packaging issue, without changing anything in the package itself. In rare cases, packages are completely broken, and are moved to a “broken” label in the conda channel, effectively removing them from being installed by default.
More often, build numbers change due to underlying dependencies across the entire Bioconda and conda-forge ecosystem. These build numbers include a hash. That hash is generated by concatenating all of the pinned versions of packages that are dependencies of that package.
For example, samtools==1.15.1=h1170115_0
refers to version 1.15.1 of
samtools
. The build number is h1170115_0
. The hash part is the
h1170115
, and the _0
refers to the first (zero-indexing) build of this
samtools version and this hash.
The hash, in turn is calculated by looking at the dependencies of samtools. The dependencies happen to include things like a C compiler (gcc), the zlib and htslib libraries and make. Some of these dependencies are “pinned”. That is, they are fixed to a particular version or versions, and those versions are used everywhere in conda-forge and Bioconda to maintain ABI compatibility (basically, to let packages co-exist in the same environment). You can find the conda-forge pinnings here, and the bioconda-specific ones here.
In the case of samtools, that hash h1170115
incorporates the packages and
versions of all of its dependencies that are pinned. That includes gcc, zlib,
and htslib. But it doesn’t include make in that hash, because make is not
pinned in those files.
The build number is likely to change, and you probably should avoid including the build number in your environment specifications – see Why shouldn’t I include build numbers in my environment YAMLs? for more information on this.
Why shouldn’t I include build numbers in my environment YAMLs?¶
As described at What’s the difference between a build number and a package version?, build numbers may change over time, independently of the actual package version. This means that when you are recording the packages installed in an environment, it is not useful to record the build number, as this is effectively over-specifying and may cause difficulty when trying to re-create the environment.
To record the installed packages in an environment, we recommend the
--no-builds
argument to conda env export
. For example, with an
environment activated:
conda env export --no-builds
The --no-builds
argument completely removes the build number from the
output, avoiding future errors when trying to rebuild the environment, and
allowing the conda solver to identify the packages that can co-exist in the
same environment.
How are dependencies pinned to particular versions?¶
In some cases a recipe may need to pin the version of a dependency.
A global set of default versions to pin against is shared with conda-forge and
can be found here.
For new dependencies that are contained in conda-forge and not yet in this list,
please update the list via a pull request.
Local pinnings can be achieved by adding a file conda_build_config.yaml
next
to your meta.yaml
.
To find out against which version you can pin a package, e.g. x.y.* or x.* please use ABI-Laboratory.
What’s the lifecycle of a bioconda package?¶
Submit a pull request with a new recipe or an updated recipe
CI (see CI Inventory) automatically builds and tests the changed recipe[s] using conda-build. Test results are shown on the PR.
If tests fail, push changes to PR until they pass.
Once tests pass, merge into master branch
CI tests again, but this time after testing the built packages are uploaded to the bioconda channel on anaconda.org.
Users can now install the package just like any other conda package with
conda install
.
Once uploaded to anaconda.org, it is our intention to never delete any old
packages. Even if a recipe in the bioconda repo is updated to a new version,
the old version will remain on anaconda.org. ContinuumIO has graciously agreed
to sponsor the storage required by the bioconda channel.
Nevertheless, it can sometimes happen that we have to mark packages as broken
in order to avoid that they are accidentally pulled by the conda solver.
In such a case it is only possible to install them by specifically considering
the broken
label, i.e.,
conda install -c conda-forge -c bioconda -c defaults -c bioconda/label/broken my-package=<broken-version>
Where can I find more info on meta.yaml
?¶
The meta.yaml
file is conda’s metadata definition file for recipes.
If you are developing a new recipe or are trying to update or improve an existing one, it can be helpful to know
which elements and values can appear in meta.yaml
.
Conda has this information available here. Please check that you are looking at the correct version of the documentation for the current conda version used by bioconda.
What are the host
and build
sections of a recipe?¶
The requirements:build
section of a meta.yaml file is used to specify the
tools for building the package, but not necessarily for running it. This is
where compilers should go. The build section might also include tools like
make
, automake
, cmake
, or git
. If there are no compilers or
other build tools, there should be no build:
section.
The requirements:host
section is used to specify shared libraries. It was
originally introduced to support cross-compiling (e.g., build on linux-64 but
output a package to be used on linux-aarch64) and the shared libraries here are
what’s needed on the target (e.g. linux-aarch64 in this example). In practice,
this is where the base interpreter python
or r-base
should go for
Python and R packages. pip
is usually here as well, and setuptools
if
it is required for the build process. cython
would go here. If a package
builds against numpy,
then numpy
should go here (otherwise, it should go in the run:
requirements). Shared libraries like zlib
, hdf5
, libcurl
, and
htslib
should go here in requirements:host
.
The requirements:run
section of a meta.yaml is used to specify the
runtime dependencies of the package.
See also
See the requirements section
of the
conda docs for more info.
Compiler tools¶
Use the syntax {{ compiler('c') }}
, {{ compiler('cxx') }}
, and/or {{
compiler('fortran') }}
. These should go in the build
section, and all
other build dependencies should go in the host
section.
Anaconda provides platform-specific compilers that are automatically
determined. The string {{ compiler('c') }}
will resolve to gcc
on
Linux, but clang
on macOS (osx-64).
See also
The
compiler tools
section of the conda docs has much more info.The default compiler options are defined by conda-build in the variants.DEFAULT_COMPILERS variable.
More details on “strong” and “weak” exports (using examples of libpng and libgcc) can be found in the
export runtime requirements
conda documentation.
How does global pinning work?¶
We can have conflicts when the version of a common library used when the package is originally built differs from the version when the package is installed. All packages intending to be installed into the same environment should be built using the same versions of common libraries so that they can co-exist. Global pinning is the idea of making sure all recipes use the same versions of common libraries.
For example, many bioinformatics tools have zlib
as a dependency.
The version of zlib
used when building the package should be the same as the
version used when installing the package into a new environment. This implies
that we need to specify the zlib
version in one place and have all recipes
intended to coexist in the same environment use that version.
This is configured with special build config files. Since we rely heavily on the conda-forge channel, the bioconda build system installs the conda-forge conda_build_config.yaml into our build environment so that it can be used for building all recipes. This is then combined with the bioconda-specific bioconda-Utils-conda_build_config.yaml. Note that in some cases the bioconda config may override some of the conda-forge configs. For example, historically, we did this when we wanted to support older Python versions.
The idea here is to specify that any time a dependency (zlib
, in our
running example) is used as a build dependency, it should also be automatically
be installed as a run dependency without having to explicitly add it as such in
the recipe. This specification is done in the zlib
recipe itself (which is
hosted by conda-forge), so in general bioconda collaborators can just add
zlib
as a build dependency.
Note that we don’t have to specify the version of zlib
in the recipe – it
is pinned in that conda_build_config.yaml
file we share with conda-forge.
In a similar fashion, the reason that we don’t have to specify libgcc
as
a run dependency is that {{ compiler('c') }}
automatically exports
libgcc
as a run dependency of any package that uses the C compiler to
build.
Understanding platform nomenclature¶
Changed on 2024-03-04: Added section
Different CPU chips use different architecture, so programs are written fundamentally differently for them. A package with compiled dependencies must therefore use platform-specific dependencies.
There is a lot of confusing nomenclature surrounding them. Here is an attempt at clearing them up, or at least providing enough context that you can look up more details on your own:
ISA, instruction set, CISC, RISC, RISC-V: The instruction set is the assembly code commands that are possible for the chip. CISC is “complex instruction set computer” which prioritizes flexibility. RISC is “reduced instruction set computer” which prioritizing power consumption (this is an oversimplification, but that’s the general idea). Instruction sets can be proprietary. Arm is a company that licenses a widely-used proprietary reduced instruction set. RISC-V is an open (non-proprietary) reduced instruction set.
Arm vs ARM: Arm is the company that licenses the proprietary instruction set. For example, they license it to Apple to run on their M-series chips. ARM (in all caps) refers to the family of RISC instruction sets, and by extension chips that use the instruction sets. It is also an acronym for Advanced RISC Machines and the eariler Acorn RISC Machine. This blog post goes into lots more detail.
x86_64
, amd64
: These are synonyms for the original Intel/AMD
architecture.
linux/x86_64
, linux/arm64
, darwin/amd64
: These are the platform
designators when using Docker (see multi-platform images in the Docker
documentation).
linux-64
, linux-aarch64
, osx-64
, osx-arm64
: These are the
platform designators used by conda in channels hosted by Anaconda.
aarch64
, arm64
: These are synonyms for ARM 64-bit architecture.
M1, M2, M3, Apple Silicon: These are chips made by Apple and used in Macs. Apple licenses the ARM RISC, so they are considered aarch64 or arm64.