[go: up one dir, main page]

Skip to content

jreades/sds_env

Repository files navigation

sds_env: Spatial Data Science Platform

This is a fork from Dani's work (please see below for citing) to remove R as we don't need this for teaching but do have a few more Python packages that we do use. We've also added some JupyterLab extensions to make interacting with the Lab server a bit easier.

We previously experimented with four approaches to installation: VirtualBox; Vagrant; Docker; and Anaconda Python directly. Each of these has pros and cons, but after careful consideration we have come to the conclusion that Docker is the most robust way to ensure a consistent experience in which all students end up with the same versions of each library, difficult-to-diagnose hardware/OS issues are minimised, and running/recovery is the most straightfoward.

=========================

For most users you should really be looking at this through the GitHub.io web site!

=========================

To Dos

  1. Work out why the Intel (AMD64) image is so much larger than the Apple Silicon (ARM64) image. I can get a decent report using docker history --format "{{.Size}}\t{{.CreatedBy}}" --no-trunc jreades/sds:2023-intel | grep -e "[G]B" which traces the difference back to just two layers:
    • 6.52GB RUN |2 USERNAME=jovyan TARGETPLATFORM=linux/amd64 /bin/bash -c mamba env update -n base --quiet --file ./${yaml_nm} && conda clean --all --yes --force-pkgs-dirs && find /opt/conda/ -follow -type f -name '*.a' -delete && find /opt/conda/ -follow -type f -name '*.pyc' -delete && find /opt/conda/ -follow -type f -name '*.js.map' -delete && pip cache purge && rm -rf /home/$NB_USER/.cache/pip && rm ./${yaml_nm} # buildkit
    • 5.48GB RUN |2 USERNAME=jovyan TARGETPLATFORM=linux/amd64 /bin/bash -c fix-permissions $CONDA_DIR && fix-permissions $HOME # buildkit
    • My guess is that the second command's effect depends on the effects of the first: there are a lot of files modified by the mamba update but they end up with a different/wrong set of permissions from what the fix-permissions script is expecting so it then has to modify the permissions on all of them which almost doubles the size of image.

Using UCL JupyterHub

Creating an Environment (Staff)

  1. Start up the UCL VPN.
  2. Connect to JupyterHub
  3. Authenticate using UCL credentials.
  4. Create a new terminal: File > New > Terminal
Incorrect Instructions from ISD

I think that these instructions are not correct (see below for the alternative) in the sense the use of a symlink can cause problems and duplicated environments down the line:

course_name="casa0013"

ln -s /shared/.../casa/${course_name} $HOME/${course_name}

conda config --add envs_dirs /shared/groups/.../casa/${course_name}/envs

curl -o /tmp/casa0013.yml https://raw.githubusercontent.com/jreades/sds_env/master/conda/environment_py.yml

conda env create -n casa0013 -f /tmp/casa0013.yml
Revised Instructions

I now think that the correct way to do this is:

course_name="casa0013"

conda config --add envs_dirs /shared/groups/.../casa/envs

curl -o /tmp/casa0013.yml https://raw.githubusercontent.com/jreades/sds_env/master/conda/environment_py.yml

conda env create -p /shared/groups/.../casa/envs -f /tmp/casa0013.yml

However, note that this now means you have .../casa/casa0013/envs/casa0013... so it might be more sensible to set envs_dirs to just ...casa/envs and then have per-module environments underneath that.

Tweaks to environyment_py.yml:

Two shortcomings in the existing approach of generating environment_py.yml were identified and need to be tweaked in the Makefile:

  1. Remove anything with ‘linux’ in it
  2. Remove SOMPY and mrmr
  3. Remove version from gitpython.
  4. Remove python-graphviz entirely.

Additional issues may exist with replication to non-Linux systems.

Connecting to an Existing Environment (PGTAs & Students)

To connect to JupyterHub:

  1. Start up the UCL VPN.
  2. Connect to JupyterHub
  3. Authenticate using UCL credentials.
  4. If you see a URL that ends in tree? please replace this with lab? to get the JupyterLab interface and not the original Jupyter Notebook interface.
  5. Create a new terminal: File > New > Terminal

Note that you need to replace ... with the appropriate path (this will be obvious logged in):

course_name="casa0013"

conda config --append envs_dirs /shared/groups/.../casa/envs

jupyter contrib nbextension install --user

Citing

This draws heavily on Dani Arribas-Bel's work for Liverpool. If you use this, you should cite him.

DOI

@software{hadoop,
  author = {{Dani Arribas-Bel}},
  title = {\texttt{gds_env}: A containerised platform for Geographic Data Science},
  url = {https://github.com/darribas/gds_env},
  version = {3.0},
  date = {2019-08-06},
}