As discussed in T93790, we should be able to drive our hardware harder while maintaining low latencies by switching to multiple Cassandra instances per hardware node.
- multiple init scripts or systemd instances
- Firejail: lighter-weight than docker, and more focus on security.
- Docker containers.
In all cases, we need a puppet-generated cassandra.yaml file per instance. Another consideration is that we normally don't want to start cassandra on boot. This is to make sure that a node that was down for a long time does not join the cluster before the config is updated.