Create a Portable Python Installation with Miniconda

Quickstart: Conda

Option A (recommended)

Build a container with Conda packages installed inside:

1. How to build your own container
2. Example container recipes for Conda
3. Use your container in your HTC jobs

Option B

Create your own portable copy of your Conda packages:

1. Follow the instructions in our guide

This approach may be sensitive to the operating system of the execution point. We recommend building a container instead, but are keeping these instructions as a backup.

More information

The above instructions are intended for if you have package(s) that need to be installed using conda install. Miniconda can be used to install Python and R and corresponding packages. But if you only need to install Python or R, and do not otherwise need to use a conda install command to set up the packages, you should see the instructions specifically for setting up Python or R because there is less chance of obscure errors when building your container.

When building or using a Miniconda container, you do not need to create or activate a conda environment. For the build process, you skip directly to the conda install commands you want to run. Similarly, when executing a script in a Miniconda container, the packages are loaded when the container starts.

Executable

If you are planning to execute a python .py script using your Miniconda container, you can follow the instructions in the Python guide.

If you are planning to execute a .R script using your Miniconda container, you can follow the instructions in the R guide.

Otherwise, you can use a bash .sh script as the submit file executable:

#!/bin/bash

<your commands go here>

where the contents of the file are the commands that you want to execute using your conda environment. You do not and should not try to activate the conda environment in the executable if you are using a container.

Specifying Exact Dependency Versions

An important part of improving reproducibility and consistency between runs is to ensure that you use the correct/expected versions of your dependencies.

When you run a command like conda install numpy, conda tries to install the most recent version of numpy. For example, numpy version 1.18.2 was released on March 17, 2020. To install exactly this version of numpy, you would run conda install numpy=1.18.2 (the same works for pip, if you replace = with ==). We recommend installing with an explicit version to make sure you have exactly the version of a package that you want. This is often called “pinning” or “locking” the version of the package.

If you want a record of what is installed in your environment, or want to reproduce your environment on another computer, conda can create a file, usually called environment.yml, that describes the exact versions of all of the packages you have installed in an environment. This file can be re-used by a different conda command to recreate that exact environment on another computer.

To create an environment.yml file from your currently-activated environment, run

[alice@submit]$ conda env export > environment.yml

This environment.yml will pin the exact version of every dependency in your environment. This can sometimes be problematic if you are moving between platforms because a package version may not be available on some other platform, causing an “unsatisfiable dependency” or “inconsistent environment” error. A much less strict pinning is

[alice@submit]$ conda env export --from-history > environment.yml

which only lists packages that you installed manually, and does not pin their versions unless you yourself pinned them during installation. If you need an intermediate solution, it is also possible to manually edit environment.yml files; see the conda environment documentation for more details about the format and what is possible. In general, exact environment specifications are simply not guaranteed to be transferable between platforms (e.g., between Windows and Linux). We strongly recommend using the strictest possible pinning available to you.

To create an environment from an environment.yml file, run

[alice@submit]$ conda env create -f environment.yml

By default, the name of the environment will be whatever the name of the source environment was; you can change the name by adding a -n <name> option to the conda env create command.

If you use a source control system like git, we recommend checking your environment.yml file into source control and making sure to recreate it when you make changes to your environment. Putting your environment under source control gives you a way to track how it changes along with your own code.

If you are developing software on your local computer for eventual use on the CHTC pool, your workflow might look like this:

1. Set up a conda environment for local development and install packages as desired (e.g., conda create -n science; conda activate science; conda install numpy).
2. Once you are ready to run on the CHTC pool, create an environment.yml file from your local environment (e.g., conda env export > environment.yml).
3. Move your environment.yml file from your local computer to the submit machine and create an environment from it (e.g., conda env create -f environment.yml), then pack it for use in your jobs, as per Create Software Package.

More information on conda environments can be found in their documentation.

Option B: Create your own portable copy

1. Create a Miniconda installation

On the submit server, download the latest Linux miniconda installer and run it.

[alice@submit]$ wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
[alice@submit]$ sh Miniconda3-latest-Linux-x86_64.sh

Accept the license agreement and default options. At the end, you can choose whether or not to “initialize Miniconda3 by running conda init?” We recommend that you enter “yes”. Once you’ve completed the installer, you’ll be prompted to restart your terminal. Log out and log back in, and conda will be ready to use to set up your software.

If you choose “no” you’ll want to save the eval command shown by the installer so that you can reactivate the Miniconda installation when needed in the future.

2. Create a conda “environment” with your software

(If you are using an environment.yml file as described later, you should instead create the environment from your environment.yml file. If you don’t have an environment.yml file to work with, follow the install instructions in this section. We recommend switching to the environment.yml method of creating environments once you understand the “manual” method presented here.)

Make sure that you’ve activated the base Miniconda environment if you haven’t already. Your prompt should look like this:

(base)[alice@submit]$ 

To create an environment, use the conda create command and then activate the environment:

(base)[alice@submit]$ conda create -n env-name
(base)[alice@submit]$ conda activate env-name

Then, run the conda install command to install the different packages and software you want to include in the installation. How this should look is often listed in the installation examples for software (e.g. Qiime2, Pytorch).

(env-name)[alice@submit]$ conda install pkg1 pkg2

Some Conda packages are only available via specific Conda channels which serve as repositories for hosting and managing packages. If Conda is unable to locate the requested packages using the example above, you may need to have Conda search other channels. More detail are available at https://docs.conda.io/projects/conda/en/latest/user-guide/concepts/channels.html.

Packages may also be installed via pip, but you should only do this when there is no conda package available.

Once everything is installed, deactivate the environment to go back to the Miniconda “base” environment.

(env-name)[alice@submit]$ conda deactivate

For example, if you wanted to create an installation with pandas and matplotlib and call the environment py-data-sci, you would use this sequence of commands:

(base)[alice@submit]$ conda create -n py-data-sci
(base)[alice@submit]$ conda activate py-data-sci
(py-data-sci)[alice@submit]$ conda install pandas matplotlib
(py-data-sci)[alice@submit]$ conda deactivate
(base)[alice@submit]$ 

More about Miniconda

See the official conda documentation for more information on creating and managing environments with conda.

3. Create Software Package

Make sure that your job’s Miniconda environment is created, but deactivated, so that you’re in the “base” Miniconda environment:

(base)[alice@submit]$ 

Then, run this command to install the conda pack tool:

(base)[alice@submit]$ conda install -c conda-forge conda-pack

Enter y when it asks you to install.

Finally, use conda pack to create a zipped tar.gz file of your environment (substitute the name of your conda environment where you see env-name), set the proper permissions for this file using chmod, and check the size of the final tarball:

(base)[alice@submit]$ conda pack -n env-name --dest-prefix='$ENVDIR'
(base)[alice@submit]$ chmod 644 env-name.tar.gz
(base)[alice@submit]$ ls -sh env-name.tar.gz

When this step finishes, you should see a file in your current directory named env-name.tar.gz

4. Check Size of Conda Environment Tar Archive

The tar archive, env-name.tar.gz, created in the previous step will be used as input for subsequent job submission. As with all job input files, you should check the size of this Conda environment file. If >100MB in size, you should NOT transfer the tar ball using transfer_input_files. Instead, you should plan to use either CHTC’s web proxy, SQUID or large data filesystem Staging. Please contact a research computing facilitators at chtc@cs.wisc.edu to determine the best option for your jobs.

More information is available at File Availability with Squid Web Proxy and Managing Large Data in HTC Jobs.

5. Create a Job Executable

The job will need to go through a few steps to use this “packed” conda environment; first, setting the PATH, then unzipping the environment, then activating it, and finally running whatever program you like. The script below is an example of what is needed (customize as indicated to match your choices above).

#!/bin/bash

# have job exit if any command returns with non-zero exit status (aka failure)
set -e

# replace env-name on the right hand side of this line with the name of your conda environment
ENVNAME=env-name
# if you need the environment directory to be named something other than the environment name, change this line
export ENVDIR=$ENVNAME

# these lines handle setting up the environment; you shouldn't have to modify them
export PATH
mkdir $ENVDIR
tar -xzf $ENVNAME.tar.gz -C $ENVDIR
. $ENVDIR/bin/activate

# modify this line to run your desired Python script and any other work you need to do
python3 hello.py

6. Submit Jobs

In your submit file, make sure to have the following:

• Your executable should be the the bash script you created in step 5.
• Remember to transfer your Python script and the environment tar.gz file via transfer_input_files. Since the tar.gz file will almost certainly be larger than 100MB, please email us about different tools for delivering the installation to your jobs, likely our SQUID web proxy.