AWS Tags as a Lightweight Datastore
One of the most common tasks in administering clusters of machines is running regular tasks on individual machines in a cluster. For example, Cassandra benefits greatly from running repairs on individual nodes to ensure data consistency. Operating manually, an ops team would maintain a document of when each machine has been, and look at that to determine which machine should be repaired next.
However, in a world where there could be hundreds of machines, this quickly becomes unscalable. Even with only a handful of machines, a team could spend a ton of energy and time simply remembering where the source of truth lives and how to decide which machine should be repaired next.
A team could write some code to pick a node and run the necessary commands. This is a bit better, but the requirement is that the command runs regularly on each node in the cluster. Without a central record to track which nodes have been repaired, there’s no way to know when and where the last repair was run, and no way to accurately predict which node will be picked next.
Enter AWS EC2 Tags. While EC2 tagging has some restrictions, it’s an ideal place to register small bits of metadata and completely replace our old school document. For our Cassandra example, it’s easy enough to use the EC2 APIs to list all machines in an Autoscale Group, or by tag, and then apply some logic to pick a machine.
Here’s some sample code that will look at the instances in an ASG, and pick one that either hasn’t been repaired, or the node repaired the longest time ago.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
The above snippet is based on the venerable boto library. Using python for interfacing with AWS allows you to use a mature, well-tested library. You can even write tests against code that uses boto with a library of mocks called moto. In short, it looks for instances that haven’t been operated on first. This is to cover the situation where you have a brand new cluster, or when the ASG group replaces nodes. Secondly, it sorts the instances by the date of the last repair ascending, which gives you the oldest node at the zeroth element of the list.
Taking this approach to managing infrastructure allows you create simple,
testable code with very few dependencies. Had
last-repair data been stored in
a local database, it would require the team to create and maintain that
database, as well as write and test additional code. The team is now free to
spend that time on other things. Without needing to worry about the integrity of
a persistent datastore, it also means that this script can reliably be run on a
very ephemeral instance. In fact, it can be run from anywhere. Scheduled Jenkins
jobs would be a good way to run this, but with the advent of
Scheduled Lamdbda tasks,
an Ops team can cut down even further on dependencies required to perform