matplotlib.org Hosting

This repo enables matplotlib.org hosting using Caddy.

Hosted repositories

At the top level, Caddy exposes the mpl-brochure-site repository with fallback to files in the matplotlib.github.com repository. The following related projects are also exposed as toplevel directories:

• basemap
• cheatsheets
• cycler
• devdocs
• governance
• matplotblog
• mpl-altair
• mpl-bench
• mpl-gui
• mpl-sphinx-theme
• mpl-third-party
• data-prototype

Ansible configuration

When running on DigitalOcean hosting, an Ansible playbook is used to configure the server with consistent settings.

Setup

Before you can run our Ansible playbooks, you need to meet the following prerequisites:

• Create a DigitalOcean API token, and pass it to the inventory generator by setting the DO_API_TOKEN environment variable.
• If you are creating a new droplet, and want to configure DNS as well, then create a CloudFlare API token, and pass it to the Ansible playbook by setting the CLOUDFLARE_TOKEN environment variable.
• Set the vault decryption password of the Ansible vaulted file with our secrets. This may be done by setting the ANSIBLE_VAULT_PASSWORD_FILE environment variable to point to a file containing the password.
• Download all the collections the playbooks depend on with the following command:

  ansible-galaxy collection install \
    --requirements-file collections/requirements.yml
  

You may wish to use direnv to set environment variables.

Running

There is currently only one playbook:

• matplotlib.org.yml, for the main matplotlib.org hosting. This playbook operates on droplets with the website tag in DigitalOcean.

Adding a new subproject

When a new repository is added to the Matplotlib organization with documentation (or an existing repository adds documentation), it will be necessary to re-configure the server to serve those files. Note, it is currently assumed that the documentation is on the gh-pages branch of the repository, and it will be served from the top-level subdirectory with the same name as the repository (similar to GitHub Pages.) There are 4 steps to achieve this:

1. Generate a secret to secure the webhook. You can follow GitHub's instructions for creating one.

2. Add repository to Ansible:

   1. Add an entry to the repos variable at the top of matplotlib.org.yml.
   2. Add the webhook secret to files/webhook_vars.yml.

3. Re-run Ansible on the playbook like below. This should clone the new repository and update the webhook handler.

4. Configure a webhook on the new repository with the following settings:

   • Payload URL of https://do.matplotlib.org/gh/<repository>
   • Content type of application/json
   • Use the secret generated in step 1
   • Trigger only on "push" events

If everything is done correctly, the GitHub webhook should have posted an initial "ping" event successfully, and documentation should be available at https://matplotlib.org/<repository>.

Provisioning a new server

Naming

We follow a simplified version of the naming scheme on this blog post:

• Servers are named <prefix>.matplotlib.org in A records, pointing to the IPv4 address of the droplet.
• Servers get a functional CNAME alias (e.g., web01.matplotlib.org) pointing to the hostname <prefix>.matplotlib.org.
• matplotlib.org is a CNAME alias of the functional CNAME of a server.

We use planets in our Solar System for the name prefix. When creating a new server, pick the next one in the list.

Initial setup

The summary of the initial setup is:

1. Create the droplet with monitoring and relevant SSH keys.
2. Assign new droplet to the matplotlib.org project and the Web firewall.
3. Add DNS entries pointing to the server on CloudFlare.
4. Grab the SSH host fingerprints.
5. Reboot.

We currently use a simple $12 droplet from DigitalOcean. You can create one from the control panel, or using the create.yml Ansible playbook:

ansible-playbook create.yml

This playbook will prompt you for 3 settings:

1. The host name of the droplet, which should follow the naming convention above.
2. The functional CNAME alias of the droplet.
3. The names of SSH keys to add to the droplet.

You may also pass these directly to Ansible as:

ansible-playbook create.yml --extra-vars "host=pluto functional=web99 ssh_keys='a b c'"

The playbook will create the server, as well as add DNS records on CloudFlare. Note, you must set DO_API_TOKEN and CLOUDFLARE_TOKEN in the environment to access these services.

Then, to ensure you are connecting to the expected server, you should grab the SSH host keys via the DigitalOcean Droplet Console:

for f in /etc/ssh/ssh_host_*_key; do
  ssh-keygen -l -f $f;
done

Note down the outputs to verify later, e.g.,

# Use these for comparison when connecting yourself.
1024 SHA256:J2sbqvhI/VszBtVvPabgxyz6sRnGLrZUn0kqfv4doAM [email protected] (DSA)
256 SHA256:J0rOMayXhL1+5wbm4WQNpAvmscDjqwJjAtk1SLemRMI [email protected] (ECDSA)
256 SHA256:y8EDRGMpLWOW72x47MVKsAfSAl8JHjsOc/RGaiMTPGs [email protected] (ED25519)
3072 SHA256:AyuNO8FES5k9vobv0Pu9XpvtjVFZ1bTTNxb1lo+AuRA [email protected] (RSA)

Finally, you should reboot the droplet. This is due to a bug in cloud-init on DigitalOcean, which generates a new machine ID after startup, causing system logs to be seem invisible.

Running Ansible

You must setup Ansible as described above. Verify that the new droplet is visible to Ansible by running:

ansible-inventory --graph

which should list the new droplet in the website tag:

@all:
  |--@website:
  |  |--venus.matplotlib.org

Then execute the Ansible playbook on the servers by running:

ansible-playbook --user root matplotlib.org.yml

During the initial "Gathering Facts" task, you will be prompted to accept the server's SSH fingerprint, which you should verify against the values found earlier. If there are existing servers that you don't want to touch, then you can use the --limit option. If you are using a non-default SSH key, you may wish to use the --private-key option.

Flip main DNS

You can verify that the server is running correctly by connecting to https://<prefix>.matplotlib.org in your browser.

Once everything is running, you should flip the DNS for the main site, changing the matplotlib.org CNAME to point to the new server's webNN.matplotlib.org functional name.