Copyright 2015 Hewlett-Packard Development Company, L.P. This work is licensed under a Creative Commons Attribution 3.0 Unported License. http://creativecommons.org/licenses/by/3.0/legalcode
Ansible Puppet Apply¶
Our current Config Management system consists of using ansible to orchestrate running puppet agent on each of our hosts, which then have to connect back to a puppetmaster to learn what they should apply. It’s a highly coupled system that ultimately doesn’t need to be that complex.
While the puppetmaster has served us well for quite some time, with the introduction of ansible for sequencing and orchestration, it’s no longer needed. Having the puppetmaster introduces a few complexities, such as the need for certificate exchange and the certificate re-use dance when we’re replacing servers. It’s also a bootstrapping issue when spinning up a new infra, since you need the puppetmaster in place to get the puppetmaster in place.
Finally, our friends from puppet sometimes like to release new versions of puppet that are backwards incompatible in ways that touch the wire protocol. If we’re using puppet apply on each of the nodes, then rolling out a per-node upgrade is simple.
Add a playbook that handles cloning system-config on all of the nodes and potentially running install_modules.sh on them. This should run before puppet is run. As an optimization, we could fetch on the master and rsync to all of the nodes, or do clarkb’s trick of fetching the ref on the master and passing that to the remote git clone/update commands.
Split the current hiera common.yaml into a set of files, a common.yaml file that contains things that should be available everywhere. A set of $group.yaml files, which map to a group of servers where every server in the group should have access to the values. And a set of $fqdn.yaml files that contain values that should only be accessed by one server.
Group membership from the puppet perspective will be managed by a group variable set at the node level in site.pp. There should be an ansible group for every defined puppet group. For the purposes of this, that grouping will be managed manually, but as a follow on, we should find a better way to manage this mapping. Also, for the purposes of this, each node will only have one group from puppet/hiera perspective. Ansible supports multiple groups, as does the ansible openstack inventory module. For now, we’ll avoid making use of that.
In the ansible-puppet role, copy over the master’s hieradata that is appropriate for that node to have. That is, the common.yaml file, any $group.yaml files for the node in question, and the $fqdn.yaml file. Then, have ansible run puppet apply pointing at the manifests/site.pp file on the host.
We will need to populate the hiera files as part of node bootstrap. This will be simple once we have replaced launch_node.py with an ansible playbook using the openstack modules. However, that work actually wants this work first to simplify the exchange, so for now, we will want a patch to launch_node.py that calls the ansible puppet role instead of doing the ssh to the host to run puppet itself.
To handle disabled nodes, we’ll mark all of our ansible playbooks in run_all with ;!disabled - referencing an ansible node group called disabled that we can add and remove hosts from on the master.
Since we will not have a puppetmaster, we’ll need to coordinate puppet apply reporting to puppetdb. For this, we need to create a CA, a cert that the puppetdb server will have, and a cert that we will distribute to every host we run. The certs for the cattle and the puppetdb server can be installed via hiera secrets, with the cattle cert being an excellent target for common.yaml.
We could just keep using puppet agent.
We could ditch puppet altogether and move completely to ansible.
- Primary assignee:
Use Gerrit topic “puppet-apply” for all patches related to this spec.
git-review -t puppet-apply
The description is kinda already written like work items.
Each server will be contacting puppetdb directly, rather than funneling through the puppetmaster.
Secrets related to a particular server will be stored on that server. Of course, those secrets are all aready on that server in some form anyway.
We already have puppet apply tests, so that’s awesome.