CoreOS Fest 2016

While etcd was born in the cloud era, it does not really play well in a dynamic environment where nodes come and go and where IP addresses are ephemeral. Moreover, etcd is meant – with its RAFT algorithm at the core – as a consistent key-value store. It rather refuses to form or join a cluster than putting consistency at risk. This talk will discuss how to make etcd more elastic in a sense that cluster join, member replacement and cluster growth is handled automatically when etcd starts up. The existing discovery service is used for this, but in contrast to the stock etcd not only during bootstrapping, but during the whole life-cycle of an etcd cluster. The ideas are implemented in the experimental elastic-etcd wrapper around etcd2. Its usage will be shown and how it behaves in certain cluster situations.