Vaidas Jablonskis bio photo

Vaidas Jablonskis

Email Twitter Google+ LinkedIn Github

My life managing systems, especially distributed, has gone much easier since I started using more and more of CoreOS awesomeness.

CoreOS is a new Linux distribution that has been rearchitected to provide features needed to run modern infrastructure stacks.

The main features of CoreOS which I like are that it is very minimal, read-only and most importantly comes with etcd out of the box.

As it stands, CoreOS is not designed to run in EC2 auto scaling group. Mostly because to create an ASG launch configuration, you have to specify which AMI id to use for launching instances. CoreOS has its own update mechanism, similar to ChromeOS and is based on Omaha spec. CoreOS cluster can be auto-updated whenever there is a new release of the channel that you are running, while ASG will always start new instances using the AMI you originally specified. This could mean that new machines which are added to the cluster can take long time to catch up with the rest of the cluster in terms of CoreOS version. Also, worth noting that they may never be able to join the cluster, because of incompatible versions of etcd (not that this is the case right now, but it might be if AMI in ASG launch configuration is very old).

If you’re using EC2 ASG, you probably doing it via CloudFormation templates. CF is very convenient, but it has its limits. One of which you have to be very careful about is CF stack updates. Each resource has a different update policy. Most dangerous are the ones which require a resource replacement if some of its properties are to be updated. In our case, it is ImageId property of AWS::AutoScaling::LaunchConfiguration type. Currently there is no way to say: hey, don’t touch currently running instances, but when you launch new ones, make sure to use an updated ImageId.

What I do is I have automatic CoreOS updates disabled by default. You can do that via user-data using cloud-config:


    reboot-strategy: off

This means, that you will have to roll out CoreOS updates as part of the CloudFormation stack update. This works for small or mid-size clusters. However, there is a catch here. As I mentioned before, CoreOS comes with etcd and I am sure you use etcd if you have more than one CoreOS node. The problem with CF stack updates and EC2 ASG in general is that it will happily carry on killing and bringing nodes one by one by default. ASG has no idea about your cluster state. Your cluster can get into a broken state very quickly, due to lack of etcd quorum, ASG replaces instances too quickly.

At etcd cluster formation time, you can set etcd member remove delay timeout, by default it is quite high for our needs. It configures etcd with the minimum time in seconds that a machine has been observed to be unresponsive before it is removed from the cluster. What we need to do is to configure ASG to wait as long as twice the amount of time it takes for etcd to notice and remove dead nodes from the cluster.

ASG has an update policy for rolling updates. Let’s set PauseTime to PT10M and MinInstancesInService to N-1, N being a desired number of nodes in your CoreOS cluster.

Again, etcd can be configured via the same cloud-config configuration using EC2 user-data:


    reboot-strategy: off
    discovery:  # <-- get a new token
    # some other etcd params
    # peer-addr etc
    cluster-remove-delay: 300

Please note, that you cannot change cluster-remove-delay parameter if your cluster is already formed.

However, I have noticed that sometimes etcd would get stuck and won’t re-elect a new leader. I have not had a chance to debug that in more detail, but I suspect that’s a bug in etcd. I am aware that etcd team is working hard at addressing these issues in version 0.5.0. One way of forcing re-election is by stoping etcd service on all CoreOS nodes and starting them again:

$ sudo systemcl stop etcd.service; sleep 5; sudo systemctl start etcd.service

I am sure that people have different ways of running CoreOS on EC2 and maybe in ASG. If you do, please share.