Hi William,

When you start a new server, please try initializing with an empty group so
that it will wait for the setConf.

Tsz-Wo

On Fri, Nov 18, 2022 at 6:38 PM William Song <[email protected]> wrote:

> Hi dev,
>
> Correctly scaling the cluster has always been a non-trivial task. There
> were some previous discussions in user mailing list, see
> https://lists.apache.org/[email protected]:2022-7:scale <
> https://lists.apache.org/[email protected]:2022-7:scale>. In
> that thread, we reached consensus as Tsz-Wo concluded as
>
> > The original design was to call setConf first and then start the nodes.
> I understand that it may not be convenient to enforce such an order. We may
> consider having an allow-list so that the Leader won't shut down the nodes
> in the allow-list.
>
> IoTDB adopted this order (setConf first and then start) in a recent PR
> https://github.com/apache/iotdb/pull/7712 <
> https://github.com/apache/iotdb/pull/7712>. However, new errors occurred,
> https://issues.apache.org/jira/browse/RATIS-1749 <
> https://issues.apache.org/jira/browse/RATIS-1749> as one example.
>
> So I investigated on Ratis users and found out that Ozone and Alluxio both
> will start server first and then call setConf. IoTDB originally takes the
> same order, although SHUTDOWN error did occurred a few.
>
> So I’m a little confused now. Is there a best practice Ratis recommends to
> scale the raft cluster? Any suggestions would be much appreciated.
>
> Best Regards,
> William

Reply via email to