Updating unicast node list without restarting

[agp]

Hi,
I have gone through a lot of threads around this topic and it seems like the answer is No, but just wanted to confirm.

Here are the details:

1. Can't use multicast
2. Most common scenario will have only 2 nodes (hence both will be master). May have more than 2 in some installations -- but lets ignore this for now.
3. Don't have list of IPs upfront. The 2 nodes may die/restart and I want to avoid having to restart the OTHER nodes to update the IP.

For ex I did this:

1. Started Node1 with unicast.hosts = Node1
2. Started Node2 with unicast.hosts = Node2 (I could do this since I knew Node1 by now).

Now if Node1 gets killed and restarted, I see it never joins the cluster even though Node1 and Node2 have the same cluster name.

The only way I understand this may work is that when Node2 is added, I should go and change Node1's unicast.hosts to be [Node1, Node2]. But this will mean that elasticsearch is being restarted on Node1 unnecessarily. Am I right?

I also read that the cluster update API ignores unicast.hosts as it is not dynamically update-able.

Linux daemons usually support -HUP signal mechanism to re-read configuration file without restarting the process. Does ES support this? If not, is there ANY way to NOT have to restart Node1 to update the unicast.hosts?

Hopefully I explained my question clearly.
Thank you.

[warkolm (Mark Walkom)]

@agp

For ex I did this:1) Started Node1 with unicast.hosts = Node12) Started Node2 with unicast.hosts = Node2 (I could do this since I knew Node1 by now).

Now if Node1 gets killed and restarted, I see it never joins the cluster even though Node1 and Node2 have the same cluster name.

Yeah, but you don't have multicast enabled so they will never see each other.

@agp

The only way I understand this may work is that when Node2 is added, I should go and change Node1's unicast.hosts to be [Node1, Node2]. But this will mean that elasticsearch is being restarted on Node1 unnecessarily. Am I right?

You should just point node 2 to node 1, there's no need to update node 1 unless you do restart it for some other reason.

@agp

the answer is No

Ultimately, ^

[anhlqn (Anh)]

I just list all master node IPs in the unicast setting on every master node. Is it ok?

[jasontedor]

@anhlqn

I just list all master node IPs in the unicast setting on every master node. Is it ok?

I assume that you meant "master node IPs and ports", and if instead you meant "on every node", yes, this is reasonable, fairly standard, and the recommended way to do it.

[anhlqn (Anh)]

Yes, that's what I have right now.

[agp]

@warkolm

You should just point node 2 to node 1, there's no need to update node 1 unless you do restart it for some other reason.

Correct. This is what I ended up doing. Whenever node1 has to restart (because it crashed or any such reason), I then update its unicast.hosts before starting elasticsearch on it. Thanks.