You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ratis.apache.org by "Siyao Meng (Jira)" <ji...@apache.org> on 2026/10/19 22:41:00 UTC
[jira] [Assigned] (RATIS-2754) Add a startup option to refresh peer addresses in the Raft configuration
[ https://issues.apache.org/jira/browse/RATIS-2754?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Siyao Meng reassigned RATIS-2754:
---------------------------------
Assignee: Siyao Meng
> Add a startup option to refresh peer addresses in the Raft configuration
> ------------------------------------------------------------------------
>
> Key: RATIS-2754
> URL: https://issues.apache.org/jira/browse/RATIS-2754
> Project: Ratis
> Issue Type: Improvement
> Reporter: Siyao Meng
> Assignee: Siyao Meng
> Priority: Major
>
> Follow up of RATIS-2082. As discussed in [https://github.com/apache/ratis/pull/1630], setConfiguration only records an address only change without rebinding or reconnecting, so a startup option is preferred.
> h3. Problem
> Peer addresses are persisted in raft-meta.conf and in the configuration entries of the Raft log. When peer addresses change (e.g. migrating IPv4 literal addresses to IPv6, or correcting a wrong address), the stale addresses are recovered on restart even if the server is started with the correct peers. The only fix today is the offline ratis-shell {{local raftMetaConf}} command, which drops listeners and priorities and requires replacing raft-meta.conf by hand on every server.
> Use case: Apache Ozone SCM HA migrating IPv4 literal Ratis peer addresses to IPv6 (HDDS-16136, under the Ozone IPv6 umbrella HDDS-15763). OM HA connections already follow its configured peers, but its stored peer addresses stay stale without this option.
> h3. Proposal
> Add a server config {{raft.server.startup.refresh-peer-addresses.enabled}} (default false). When it is enabled and a group is started with a non-empty list of peers, the server refreshes the stored peer addresses at startup, following the sequence in [this comment|https://issues.apache.org/jira/browse/RATIS-2754?focusedCommentId=18127578&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-18127578]:
> # Start as usual, but do not take server or client requests yet.
> # Apply the log up to the commit index recovered from the log metadata ({{raft.server.log.log-metadata.enabled}}, default true). Only these entries are known to be committed at startup.
> # Write the last applied conf to raft-meta.conf at the same index, with the addresses (address, admin, client and data stream) of the startup peers replacing those of the peers with the same ids. Membership, roles, listeners and priorities are kept. Conf entries beyond the commit index get the new addresses in memory only.
> # Start taking requests.
> Rolling restart: a peer that has not been restarted still has the old addresses. If it becomes the leader, its startup conf entry brings the old addresses back. So a restarted peer remembers the addresses it replaced at startup and replaces them again whenever a later conf carries them, both in memory and when writing raft-meta.conf. Other address changes (e.g. via setConfiguration) are kept. A rolling restart then converges in any order.
> A startup peer not in the configuration is ignored with a warning. Membership changes still go through setConfiguration.
> Changing peer addresses then becomes: update the startup peers, enable the option and do a rolling restart. Keep the option enabled until all the servers have been restarted with it.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)