Why bother with masterless?
We had two different use cases at my work that led to my writing Miyamoto. First, we manage our Macintoshes with puppet, and we have users in remote locations that don't always need to connect to the VPN. When their machines aren't on the VPN or corp network, they can't connect to the puppetmaster and don't receive updates. Miyamoto allows them to get their manifests from an S3 bucket, so as long as they have internet, they get puppet updates.
Secondly, we have some compute-intensive tasks that we do in EC2. As the load on the cluster rises and falls, we add and remove instances. Hundreds of thousands of instances over time. Way more than I care to have to deal with when cleaning stale certs out of the ca certs directory. By using an S3 bucket, we eliminate the SPOF of a single puppetmaster, or the maintenance hassle of a cluster of puppetmasters and ca boxes.
But we're losing reporting
Miyamoto writes the facts collected by facter to a status file in the S3 bucket, so you still can see what is happening with our fleet by scraping those files and loading them into your reporting system of choice. They're in json format so they're easy to parse.
If you're going to go masterless for both your macs and your servers, I recommend that you use different directory trees in S3. I use
bucket/ubuntu/binaries bucket/ubuntu/manifests bucket/ubuntu/status
for my servers, and
bucket/osx/binaries bucket/osx/manifests bucket/osx/status
for my OS X machines.
I use a different AWS user for each architecture tree, and separate AWS groups to control the acls for the trees. The machine users only have write permission on the status subdirectory of their architecture subtree, and read-only access to the rest of their architecture tree. A compromised machine can't push new manifests to its own architecture, and it can't view the manifests or statuses of the other architectures.
I've included an example AWS policy that enforces this. It assumes that you're dedicating an entire AWS bucket to your masterless setup, and I recommend that you do so so you don't have to worry about problems in the policy leaving the rest of the files in your bucket vulnerable to attack if someone gets their hands on one of your masterless machines.
- AWS account
- Tim Kay's aws script (get the latest version from http://timkay.com/aws/)
- Dedicated S3 bucket
If you're planning to administer Debian/Ubuntu clients, you'll need Jordan Sissel's fpm gem. If you're administering OS X clients, you'll need to install my Luggage tool (http://github.com/unixorn/luggage).
If you're planning to use this with RedHat/Centos, it shouldn't take much tweaking to have fpm spit out rpms instead of debs and have mm_update_manifests cope accordingly.
Fun with Rake
The Rakefiles assume that you are going to have a stable and a development environment, and that they'll be in separate git branches. Set stable in config.json to the name of your stable branch, and development to your development environment.
rake -T for all the available rake tasks
- Create an S3 bucket (masterless-puppet.example.com)
- Create the directory tree you want in your bucket as shown in the Base Setup
- Create an AWS user for your machines with iam-usercreate. If you're puppeting more than one architecture, I recommend one user per architecture, and keep the architectures in separate repositories
- Create an AWS group for each architecture with iam-groupcreate
- Modify the enclosed policy to use your bucket and directory tree, then apply it to your AWS group
- Copy base_config.json to config.json, edit appropriately for your setup
- sudo rake experimental_checkout will create a deb of the current configuration without changing branches