Switch environments before running Jupyter kernels

Sometimes, one needs to execute Jupyter kernels in a different environment. Say you want to execute the kernel in a conda environment (that's easy, but actually misses setting certain environment variables). Or run it inside a Docker container. One could manually adjust the kernelspec files to set environment variables or run commands before starting the kernel, but envkernel automates this process.

envkernel is equally usable for end users (on their own systems or clusters) to easily access environments in Jupyter, or sysadmins deploying this access on systems they administer.

In general, there are two passes: First, install the kernel, e.g.: envkernel virtualenv --name=my-venv /path/to/venv. This parses some options and writes a kernelspec file with the the --name you specify. When Jupyter tries to start this kernel, it will execute the next phase. When Jupyter tries to run the kernel, the kernelspec file will re-execute envkernel in the run mode, which does whatever is needed to set up the environment (in this case, sets PATH to the /path/to/venv/bin/ that is needed). Then it starts the normal IPython kernel.

Available modes:

• conda: Activate a conda environment first.
• virtualenv: Activate a virtualenv first.
• docker: Run the kernel in a Docker container.
• singularity: Run the kernel in a singularity container.
• Lmod: Activate Lmod modules first.

Installation

Available on the PiPI: pip install envkernel.

Or, you can install latest from Github in the usual way: pip install https://github.com/NordicHPC/envkernel/archive/master.zip

This is a single-file script and can be copied directly and added to PATH as well. By design, there are no dependencies except the basic Jupyter client (not notebook or any UI), and that is only needed at kernel-setup time, not at kernel-runtime. The script must be available both when a kernel is set up, and each time the kernel is started (and currently assumes they are in the same location).

General usage and common arguments

General invocation:

envkernel [mode] [envkernel options] [mode-specific-options]

General arguments usable by all classes during the setup phase:

These options directly map to normal Jupyter kernel install options:

• mode: singularity, docker, lmod, or whatever mode is desired.
• --name $name: Name of kernel to install (required).
• --user: Install kernel into user directory.
• --sys-prefix: Install to the current Python's sys.prefix (the Python which is running envkernel).
• --prefix: same as normal kernel install option.
• --display-name NAME: Human-readable name.
• --replace: Replace existing kernel (Jupyter option, unsure what this means).
• --language: What language to tag this kernel (default python).

These are envkernel-specific options:

• --verbose, -v: Print more debugging information when installing the kernel. It is always in verbose mode when actually running the kernel.
• --python: Python interpreter to use when invoking inside the environment. (Default python. Unlike other kernels, this defaults to a relative path because the point of envkernel is to set up PATH properly.) If this is the special value SELF, this will be replaced with the value of sys.executable of the Python running envkernel.
• --kernel=NAME: Auto-set --language and --kernel-cmd to that needed for these well-known kernels. Options include ipykernel (the default), ir, or imatlab. But all of these hard-code a kernel command line and could possibly be wrong some day.
• --kernel-cmd: a string which is the kernel to start - space separated, no shell quoting, it will be split when saving. The default is python -m ipykernel_launcher -f {connection_file}, which is suitable for IPython. For example, to start an R kernel in the environment use R --slave -e IRkernel::main() --args {connection_file} as the value to this, being careful with quoting the spaces only once. To find what the strings should be, copy form some existing kernels. --kernel=NAME includes shortcut for some popular kernels.
• --kernel-template: An already-installed kernel name which is used as a template for the new envkernel. This is searched using the normal Jupyter search paths. This kernel json file is loaded and used as a template for all kernel options (--language, --kernel-cmd, etc). Also, any other file in this directory (such as logos) are copied to the new kernel (like kernel.js in irkernel).
• --kernel-make-path-relative removes an absolute path from the kernel command (mainly useful with --kernel-template). This would be useful, for example, where you are setting up an lmod install and the absolute path of the module might change, but you want it to always run Python relative to that module anyway.
• --env=NAME=VALUE. Set these environment variables when running the kernel. These are actually just saved in the kernel.json file under the env key, which is used by Jupyter itself. So, this is just a shorthand for adding variables there, it is not used at the envkernel stage at all.

Order of precedence of options (later in the list overrides earlier): --kernel-template, --kernel, --kernel-cmd, --language, --python, --display-name.

Conda

The Conda envkernel will activate Conda environments (set the PATH, CPATH, LD_LIBRARY_PATH, and LIBRARY_PATH environment variables). This is done manually, if anyone knows a better way to do this, please inform us.

Conda example

This will load the anaconda environment before invoking an IPython kernel using the name python, which will presumably be the one inside the anaconda3 environment.

envkernel conda --name=conda-anaconda3 /path/to/anaconda3

Conda mode arguments

General invocation:

envkernel conda --name=NAME [envkernel options] conda-env-full-path

• conda-env-full-path: Full path to the conda environment to load.

Virtualenv

This operates identically to conda mode, but with name virtualenv on virtualenvs.

Virtualenv example

envkernel virtualenv --name=conda-anaconda3 /path/to/anaconda3

Docker

Docker is a containerization system that runs as a system service.

Note: docker has not been fully tested, but has been reported to work.

Docker example

envkernel docker --name=NAME  --pwd --bind /m/jh/coursedata/:/coursedata /path/to/image.simg

Docker mode arguments

General invocation:

envkernel docker --name=NAME [envkernel options] [docker options] [image]

• image: Required positional argument: name of docker image to run.

• --pwd: Bind-mount the current working directory and use it as the current working directory inside the notebook. This is usually useful.

• A few more yet-undocumented and untested arguments...

Any unknown argument is passed directly to the docker run call, and thus can be any normal Docker argument. If ,copy is included in the --mount command options, the directory will be copied before mounting. This may be useful if the directory is on a network mount which the root docker can't access. It is recommended to always use the form of options with =, such as --option=X, rather than separating them with a space, to avoid problems with argument/option detection.

Singularity

Singularity is a containerization system somewhat similar to Docker, but designed for user-mode usage without root, and with a mindset of using user software instead of system services.

Singularity example

envkernel singularity --name=NAME --contain --bind /m/jh/coursedata/:/coursedata /path/to/image.simg

Singularity mode arguments

General invocation:

envkernel singularity --name=NAME [envkernel options] [singularity options] [image]

• image: Required positional argument: name of singularity image to run.

• --pwd: Bind-mount the current working directory and use it as the current working directory inside the notebook. This may happen by default if you don't --contain.

Any unknown argument is passed directly to the singularity exec call, and thus can be any normal Singularity arguments. It is recommended to always use the form of options with =, such as --bind=X, rather than separating them with a space, to avoid problems with argument/option detection. The most useful Singularity options are (nothing envkernel specific here):

• --contain or -c: Don't share any filesystems by default.

• --bind src:dest[:ro]: Bind mount src from the host to dest in the container. :ro is optional, and defaults to rw.

• --cleanenv: Clean all environment before executing.

• --net or -n: Run in new network namespace. This does NOT work with Jupyter kernels, because localhost must currently be shared. So don't use this unless we create proper net gateway.

Lmod

The Lmod envkernel will load/unload Lmod modules before running a normal IPython kernel.

Using envkernel is better than the naive (but functional) method of modifying a kernel to invoke a particular Python binary, because that will invoke the right Python interpreter but not set relevant other environment variables (so, for example, subprocesses won't be in the right environment).

Lmod example

This will run module purge and then module load anaconda3 before invoking an IPython kernel using the name python, which will presumably be the one inside the anaconda3 environment.

envkernel lmod --name=anaconda3 --purge anaconda3

Lmod mode arguments

General invocation:

envkernel lmod --name=N