posted 2012-06-26 2001 by masen
Working at ResTek over the summer is a much different experience than
working during the school year. For one: I am now working 40 hours per week (up from
15), and also many more people are working in the office regularly which makes
collaboration easier. All this leads to an environment where things get done in a
third or less of the time it would 'normally' take to complete. And the bugfixing is
The first big project that's nearing completion is a new deployment of office workstations for staff to use. We have about 17 machines which dual boot between Windows 7 Enterprise (deployed with Symantec Ghost) and Lubuntu 12.04 (deployed with puppet 2.7). We have a much more uniform configuration than we previously did, but there are still many faults to discover and fix.
I feel like everything about this story comes back to DNS because we as people do not like to remember numbers, especially not ones which may change on a daily basis. That's why DNS was created! So the goal is to have puppet periodically refresh the configuration on each host, and in the process notify the server of its current IP address. We can later use this information to feed the DNS server and have up-to-date dynamic dns for all machines in our office. We deployed the manifest in a simple way that makes a GET request to the puppetmaster server with the client's hostname. A PHP script then aggregates that information into the database. Surprisingly, this part of the system worked just fine. After pushing the change and manually refreshing puppet, I was able to get all the hosts to report their IP address every 30 minutes...Or so I thought.
Upon arriving the next day, I noticed that some of the computers were no longer refreshing their configuration from the puppetmaster, or the script to update their IP was not working. After the checking the log on an affected server, I saw a message like this:
Jun 21 08:52:47 Kiwi-Strawberry puppet-agent: Did not receive certificate
Repeating over and over in the logs, every 2 minutes. It had to be a puppet issue. I restarted the puppet agent...and it refreshed normally. After restarting the agent, it continued to work just fine. Meanwhile, I continued to puzzle about it and search google, only to uncover 3 year old posts and other worthless information. I was going to have to solve this one myself.
After these seemingly random puppet failures continued to happen over the next few days on different servers I began to notice a pattern: machines that had been restarted didn't work until the puppet service was restarted separately and later. A freshly booted machine will sit there and wait for a certificate despite the fact that it already has one signed and working! I could now reproduce the issue and started looking for other clues.
I noticed on the puppetmaster that several certificates were in the signing cache waiting to be signed. They corresponded to the bare (non-FQDN) hostname of each client which was experiencing post reboot issues. It started to click in my mind.
I restarted the workstation and then fired up my editor to look at the logs. Sure enough, the puppet-agent (which is started via a sysv init.d script on Ubuntu) was starting BEFORE the network was up. This was causing puppet to see a different hostname (without the DHCP supplied domain suffix) and consequently generate a different certificate. When puppet finally got network access it was requesting that the puppetmaster sign its new certificate; of course this is not the right thing and our site.pp manifest is not set up to manage nodes without a fully qualified domain name.
Several solutions were proposed to tackle this problem from: having cron restart puppet occasionally to renaming init.d scripts. However, nothing seemed quite as nice as moving to an event-based startup for puppet. It was easy enough to hack together an upstart configuration file for puppet:
respawn console log start on (local-filesystems and net-device-up IFACE=eth0) stop on runlevel [!12345] pre-start script test -z $UPSTART_EVENTS || echo Puppet started by: $UPSTART_EVENTS >> /var/log/puppetd.log end script exec /usr/bin/puppet agent --no-daemonizeAfter placing this in /etc/init/puppet.conf, I disabled the init.d startup script:
# chmod 644 /etc/init.d/puppetOn the next reboot, puppet waited until the network was fully up and communication active before being started. The problem was resolved! I quickly added the solution to our office manifest and pushed the changes.