Clarification on using --initial-cluster-state=new for multi-site etcd deployment

[prateekkohli21]

Hi,

We are deploying an etcd cluster across multiple Kubernetes clusters (multi-site). Each site runs one etcd member, and all members are defined statically using:

--initial-cluster=
--initial-cluster-state=new

We observed that:
• Even if one site is deployed first and others start minutes later,
• And even if all members are started with --initial-cluster-state=new,
• The cluster forms correctly and members join successfully.

This seems to work reliably in our tests.

However, based on documentation our understanding is:
• new should only be used during initial bootstrap,
• Subsequent members should use existing,

Questions:
1. Is it officially supported to start all members with --initial-cluster-state=new if all members are defined in --initial-cluster and we start them 1 by 1?
2. Is using existing strictly required after the first member starts, or only when dynamically adding members?

We want to ensure this deployment model is safe and supported.

Thanks!!

[ahrtr]

1. Is it officially supported to start all members with --initial-cluster-state=new if all members are defined in --initial-cluster and we start them 1 by 1?

YES

2. Is using existing strictly required after the first member starts, or only when dynamically adding members?

"existing" is only required when you start a newly added member. For example, you just add a new member by executing etcdctl member add xxx, and when you start the new member, you must use existing in such case.

When you start the cluster in the very first time and there isn't any local data yet, then you must use new.

If there is already local data, then it doesn't matter what's the value. (of course, it's recommended to use existing in such case).

Also refer to https://github.com/ahrtr/etcd-issues/blob/master/docs/cluster_id_mismatch.md