Etcd v0.1.2 with a new dashboard
coreos.com
coreos.com
Right now we have a service, called coordination (c10n), which helps boot strap an example cluster on ec2. The basic idea is that in order to get things going you need to know the IP of at least one other machine and c10n holds on to this information. The root for this bootstrap is a unique token you put in the AWS user data.
You either need an external service like c10n with a shared token or a stable hostname that you setup yourself that can distribute the initial seed of information. An external service makes for an easier experience.
> and how do these server know to trust it?
As we get further along we will be adding a API/dashboard to help with boot strapping on etcd and approving/denying new machines as followers. That is part of why we added client certs to etcd from the very beginning.
Re: autoscaling: Our focus at the moment is getting the core services, like etcd, working properly so that we can correctly build higher order concepts like horizontal auto scaling. More to come. Thanks for the questions.
Also, all the examples just store simple key/value pairs. What are your best practices for actual config files? Do you store e.g. key=/etc/nginx/nginx.conf value=<contents of that file>? Does that imply all machines will have the same config? Or do you have the machine name in the key space? Or is etcd not actually for this kind of configs? If not, for which kind then?
You build master election on top of the test and set functionality of etcd which gives you a cluster wide locking mechanism. The plan is to provide a simplified master election REST api as a module, just like the dashboard in this release is a module.
But, nothing is stopping a regular etcd client from building master election on top of the current API.
Something I would like to add to etcdctl is a config file generator which reads from the keys of etcd and fills in fields. You could add the entire file but it is more reasonable to simply add the metadata that is dynamic into etcd and ship a template around. Another example of this is this chef integration: https://github.com/coderanger/etcd-chef
For some applications, like load balancers, you may want to have the process itself interacting with an etcd cluster. e.g. this HTTP load balancer prototype: https://github.com/calavera/active-proxy
Could you give some pointers on how to implement master election on top of test-and-set? What is the algorithm? I've tried searching for it, but I've likely missed correct keywords.
[1] http://redis.io [2] http://highscalability.com/blog/2013/8/7/raft-in-search-of-a... [3] http://labs.yahoo.com/files/ladis08.pdf
One question/problem: I can't find any kind of permissions/separation/ACLs for it. Yes, you can do authentication, but w/o even basic POSIX style owner/group/world type permissions, this essentially becomes single-user, with all apps/programs/etc trusted and able to break everyone else.
Or am I missing something?