[
https://issues.apache.org/jira/browse/SOLR-14708?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17211773#comment-17211773
]
Marcus Eagan commented on SOLR-14708:
-------------------------------------
That’s right, users may experience issues with the cache. Maybe we should be
explicit about cache as the potential cause of issues?
> Backward-Compatible Replication
> -------------------------------
>
> Key: SOLR-14708
> URL: https://issues.apache.org/jira/browse/SOLR-14708
> Project: Solr
> Issue Type: Bug
> Components: SolrCloud
> Reporter: Marcus Eagan
> Priority: Critical
>
> In [SOLR-14702|https://issues.apache.org/jira/browse/SOLR-14702] I proposed
> that we remove master/slave terminology from the Solr codebase. Now that's
> complete, we need to ensure it is backward compatible to support rolling
> upgrades from 8.7.x to 9.x because we really ought not to make it harder to
> upgrade Solr.
> Tomas offered a helpful path in a now abandoned PR:
> {quote}One way to get back compatibility and rolling upgrades could be to
> make 9.x code be able to read previous formats, but write new format, and
> make 8.x (since 8.7) read new and old, but write old? Anyone wanting to do a
> rolling upgrade to 9 would have to be on at least 8.7. Rolling upgrades to
> 8.7 would still work.
> All the code other than the requests/responses could be changed in 8_x
> branch, in addition to master.
> {quote}
> The approach that we will take is to add a ternary operator in 9_X to accept
> parameter values for the legacy verbiage, or leader/follower, but only write
> leader/follower. We need to then make 8_x work in the inverse way. The burden
> here is not on that proposal or on the code in my view. Instead, the burden
> is on the test plan.
> If anyone has any guidance please share but here are my thoughts:
> Case A:
> Test the case where a user is running a standalone cluster in 8 with three
> nodes but then updates one of the nodes.
> Case B:
> Test the case where a user is running a mixed cluster standalone cluster, and
> the leader node is forced to fail and then is brought back.
> Case C:
> A SolrCloud cluster that has a mix of 8 and 9 nodes goes down during a
> rolling upgrade and a follower needs to become leader.
> I know haven't listed all possible scenarios or everything that could happen.
> Please let me know if you have thoughts or guidance on how best to accomplish
> this work.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
