It's not a secret that I was playing/experimenting with OpenStack in the last days. When I mention OpenStack, I should even say RDO , as it's RPM packaged, built and tested on CentOS infra.
Now that it's time to deploy it in Production, that's when you should have a deeper look at how to proceed and which tool to use. Sure, Packstack can help you setting up a quick PoC but after some discussions with people hanging around in the #rdo irc channel on freenode, it seems that almost everybody agreed on the fact that it's not the kind of tool you want to use for a proper deploy.
Let's so have a look at the available options. While I really like/prefer Ansible, we (CentOS Project) still use puppet as our Configuration Management tool, and itself using Foreman as the ENC. So let's see both options.
- Ansible : Lot of natives modules exist to manage an existing/already deployed openstack cloud, but nothing really that can help setting up one from scratch. OTOH it's true that Openstack Ansible exists, but that will setup openstack components into LXC containers, and wasn't really comfortable with the whole idea (YMMV)
- Puppet : Lot of puppet modules so you can automatically reuse/import those into your existing puppet setup, and seems to be the prefered method when discussing with people in #rdo (when not using TripleO though)
So, after some analysis, and despite the fact that I really prefer Ansible over Puppet, I decided (so that it could still make sense in our infra) to go the "puppet modules way". That was the beginning of a journey, where I saw a lot of Yaks to shave too.
It started with me trying to "just" reuse and adapt some existing modules I found. Wrong. And it's even fun because it's one of my mantras : "Don't try to automate what you can't understand from scratch" (And I fully agree with Matthias' thought on this ).
So one can just read all the openstack puppet modules, and then try to understand how to assemble them together to build a cloud. But I remembered that Packstack itself is puppet driven. So I just decided to have a look at what it was generating and start from that to write my own module from scratch. How to proceed ? Easy : on a VM, just install packstack, generate answer file, "salt" it your needs, and generate the manifests :
yum install -y centos-release-openstack-ocata && yum install openstack-packstack -y
packstack --gen-answer-file=answers.txt
vim answers.txt
packstack --answer-file=answers.txt --dry-run
* The installation log file is available at: /var/tmp/packstack/20170508-101433-49cCcj/openstack-setup.log
* The generated manifests are available at: /var/tmp/packstack/20170508-101433-49cCcj/manifests
So now we can have a look at all the generated manifests and start from scratch our own, reimporting all the needed openstack puppet modules, and that's what I did .. but started to encounter some issues. The first one was that the puppet version we were using was 3.6.2 (everywhere on every release/arch we support, so centos 6 and 7, and x86_64,i386,aarch64,ppc64,ppc64le).
One of the openstack component is RabbitMQ but openstack modules rely on the puppetlabs module to deploy/manage it. You'll see a lot of those external modules being called/needed by openstack puppet. The first thing that I had to do was investigating our own modules as some are the same name, but not coming from puppetlabs/forge, so instead of analyzing all those, I moved everything RDO related to a different environment so that it wouldn't conflict with some our our existing modules. Back now to the RabbitMQ one : puppet errors where trying to just use it. First yak to shave : updating the whole CentOS infra puppet to higher version because of a puppet bug. Let's so rebuild puppet for centos 6/7 and with a higher version on CBS
That means of course testing our own modules, on our Test Foreman/puppetmasterd instance first, and as upgraded worked, I applied it everywhere. Good, so let's jump to the next yak.
After the rabbitmq issue was solved, I encountered other ones coming from openstack puppet modules now, as the .rb ruby code used for type/provider was expecting ruby2 and not 1.8.3, which was the one available on our puppetmasterd (yeah, our Foreman was on a CentOS 6 node) so another yak to shave : migrating our Foreman instance from CentOS 6 to a new CentOS 7 node. Basically installing a CentOS 7 node with the same Foreman version running on CentOS 6 node, and then following procedure, but then, again, time lost to test update/upgrade and also all other modules, etc (One can see why I prefer agentless cfgmgmt).
Finally I found that some of the openstack puppet modules aren't touching the whole config. Let me explain why. In Openstack Ocata, some things are mandatory, like the Placement API, but despite all the classes being applied, I had some issues to have it to run correctly when deploying an instance. It's true that I initially had a bug in my puppet code for the user/password to use to configure the rabbitmq settings, but it was solved and also applied correctly in /etc/nova/nova.conf (setting "transport_url=") . But openstack nova services (all nova-*.log files btw) were always saying that credentials given were refused by rabbitmq, while tested manually)
After having verified in the rabbitmq logs, I saw that despite what was configured in nova.conf, services were still trying to use the wrong user/pass to connect to rabbitmq. Strange as ::nova::cell_v2::simple_setup was included and was supposed also to use the transport_url declared at the nova.conf level (and so configured by ::nova) . That's how I discovered that something "ugly" happened : in fact even if you modify nova.conf, it stores some settings in the mysql DB, and you can see those (so the "wrong" ones in my case) with :
nova-manage cell_v2 list_cells --debug
Something to keep in mind, as for initial deployment, if your rabbitmq user/pass needs to be changed, and despite the fact that puppet will not complain, it will only update the conf file, but not the settings imported first by puppet in the DB (table nova_api.cell_mapping if you're interested) After that, everything was then running, and reinstalled/reprovisioned multiple times my test nodes to apply the puppet module/manifests from puppetmasterd to confirm.
That was quite a journey, but it's probably only the beginning but it's a good start. Now to investigate other option for cinder/glance as it seems Gluster was deprecated and I'd like to know hy.
Hope this helps if you need to bootstrap openstack with puppet !