reused cluster names break services.
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
elasticsearch (Juju Charms Collection) |
Fix Released
|
Undecided
|
Michael Nelson |
Bug Description
I have assisted in fixing two prodstack outages that were caused by re-used cluster names. The fault happens when a service is deployed with a cluster name that matches another service. The fault also happens in the reverse, a deployment starts unique, but latter a service is deployed with the cluster name, breaking a service that has not changed.
The cluster name is configurable so that someone can manually deploy a new ES and add it to an existing cluster. It just works because ES discovers clusters on the local network.
We are using juju though. the cluster is implied by the service relation. The service is automated. I propose that the charm generates a UUID to discourage devops from configuring it, or using a name that can easily be reused on a network.
This is no longer an issue with the trusty charm, which (a) uses unicast discovery based on the peer service relation, and (b) firewalls 2300 so that it's open only to peers of the service.