initial implementation of persistence [RFC/WIP]

[rmoriz]

(This is currently a RFC which I just hacked together.)

Motivation:

In complex setups, when you have more than one front-facing httpd, like in anycast/cdn setups, you have the following problems with acme workflows:

• initial bootstrap needs certificates (self-signed, no problem)
• only one node should start the acme verification process
• acme/LE needs to be directed to the right node every time
• additional node bootstrapping and renewal needs a way to distribute the key/cert

I've seen many suggestions like using s3, syncthing, rsync… however I don't like the complexity and more single points of failure.

My implementation

• uses chef's data bags, with or without encryption (strongly preferred)
• needs a chef-server or chef-zero setup (knife-zero works really well)

Limitation

The master client needs the correct permission to create/upload data bags to the chef server. This may have security implications.

One must configure one and only one node which acts as master. This can be a simple node attribute, or some DSL method like tagged?. fail-over must be implemented manually or by automatically updating attributes on the chef-server by a 3rd party (e.g. monitoring).

Key/Cert on workers may lag one converge cycle, this can be reduced by triggering converges on workers with tools like chef's "push jobs".

Usage:

Chef

Example recipe code:

# action :load checks for an existing data bag item and renders the files
#
acme_persistence 'example.com' do
  master node['acme']['master']  # this should be true only on the master node
  data_bag_name 'acme-certificates'
  alt_names %w(www.example.com example.com)
  key       '/etc/ssl/private/example.com.key'
  fullchain '/etc/ssl/certs/example.com.crt'
  action :load

  notifies :reload, 'service[nginx]'
end

acme_selfsigned 'example.com' do
  key '/etc/ssl/private/example.com.key'
  crt '/etc/ssl/certs/example.com.crt'

  not_if  'test -f /etc/ssl/certs/example.com.key && test -f /etc/ssl/certs/example.com.crt'
end

(configure nginx)

acme_certificate 'example.com' do
  alt_names %w(www.example.com example.com)
  key      '/etc/ssl/private/example.com.key'
  fullchain '/etc/ssl/certs/example.com.crt'
  method   'http'
  wwwroot  '/var/www/acme'

  only_if { node['acme']['master'] }
  notifies :save,   'acme_persistence[example.com]', :immediately # this will trigger the data bag item upload
  notifies :reload, 'service[nginx]', :immediately
end

on the master node with node['acme']['master'] set to true

first run

• no data bag item will be found
• acme_selfsigned will create a selfsigned certificate
• acme_certificate will do the acme work and trigger save to acme_persistence.
• acme_persistence uploads data to server.

subsequent runs/ renew

• data bag data found but will only be written if existing cert is self-signed or older
• acme_certificate will check expiration, re-keying, and trigger save to acme_persistence
• acme_persistence uploads data to server.

on workers, without node['acme']['master']

first run

• acme_persistence does not find certificates in the data bag
• acme_selfsigned will create dummies so the webserver can start.

subsequent runs/ renew

• acme_persistence finds a data bag item and renders out the data and triggers a webserver reload (same logic as in the master)

Webserver

Workers must proxy acme verification requests to the .well-known URL to the master node. This can easily be scripted inside a virtual host template and triggered by the same source as above (e.g. a node attribute)

e.g.

  # /etc/nginx/conf.d/upstream.conf
  upstream acme_master {
    server unique-address.webserver.example.com:443;
  }

 # vhost-template.conf.erb
...

location ^~ /.well-known/acme-challenge/ {
  default_type "text/plain";

  <% if @acme_master %>
  root /var/www/acme;
  <% else %>
  proxy_pass https://acme_master;
  <% end %>
}

location = /.well-known/acme-challenge/ {
  return 404;
}
...

related: https://github.com/schubergphilis/chef-acme/issues/15

[thoutenbos]

Nice one!

Would it make sense to be compatible with the certificates cookbook in data bag storage? Then this cookbook wouldn't even need to run on any non-master nodes.

[rmoriz]

renamed data bag item key "crt" to "cert" which should make it compatible if the combination of "key", "crt"/"cert" and "chain" is used.

[thoutenbos]

So this is definitely a great start, just thinking what we would need before it can be merged in. Of the top if my head I'm thinking:

• Update the readme
• add Kitchen integration tests
• add ChefSpec matchers

[rmoriz]

@thoutenbos can you add a basic chefspec setup to the master branch? kitchen integration tests are IMHO worthless because they fail all the time and have such a big overhead.

[rmoriz]

… and we need chef13 compatibility.

[rmoriz]

changes in SAN need to be covered, too. This is broken in the main part, too (see #35).

Reagarding SAN changes: IMHO this should be part of the core acme_certificate resource and not be part of the persistence. Persistence should restore whatever is possible, maybe even without using this cookbook. It's the job of the acme_certificate resource to renew and rerun the process on changes in SAN list (I guess only when additional/new SANs are requested).

[thoutenbos]

Now that we have Chef 13 support and the build fixes this should be able to go to master. Would you be able to add the README and integration tests to make it ready ?