[
https://issues.apache.org/jira/browse/KUDU-2247?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Alexey Serbin reassigned KUDU-2247:
-----------------------------------
Assignee: (was: Alexey Serbin)
> Update doc on the 'kudu tablet change_config move_replica' after 3-4-3
> enabled by default
> -----------------------------------------------------------------------------------------
>
> Key: KUDU-2247
> URL: https://issues.apache.org/jira/browse/KUDU-2247
> Project: Kudu
> Issue Type: Task
> Components: consensus, documentation, ops-tooling
> Affects Versions: 1.7.0
> Reporter: Alexey Serbin
> Priority: Major
>
> The way how the replica replacement works in 3-4-3 v1 design scheme has a few
> corner cases for very specific run-time scenarios. That's due to the absence
> of the {{SUPERCEDES}} attribute, which is in the full 3-4-3 design proposal
> (option E), but not in 3-4-3 v1 design.
> * Initial configuration is {{[ A(V:\+\:), B(V:\+\:), C(V:\+\:) ]}}
> * A voter replica {{A}} marked with the {{REPLACE}} attribute: {{[
> A(V:\+:REPLACE=true), B(V:\+\:), C(V:\+\:) ]}}
> * A non-voter replica {{X}} is added to replace replica {{A}}: {{[
> A(V:\+:REPLACE=true), B(V:\+\:), C(V:\+\:), X(N:\+:PROMOTE=true) ]}}
> * Replica {{B}} fails, so the system adds another non-voter replica {{Y}} to
> replace the failed replica: {{[ A(V:\+:REPLACE=true), B(V:\-\:), C(V:\+\:),
> X(N:\+:PROMOTE=true), Y(N:\+:PROMOTE=true) ]}}
> * After some time, before replica tablet copying is complete for either of
> two replicas {{X}} or {{Y}}, replica {{B}} is back, so the system evicts
> replica {{X}}: {{[ A(V:\+:REPLACE=true), B(V:\+\:), C(V:\+\:),
> Y(N:\+:PROMOTE=true) ]}}
> * Eventually, replica {{Y}} completes copying the data, catches up with the
> leader and is promoted by the leader replica: {{[ A(V:\+:REPLACE=true),
> B(V:+\:), C(V:\+\:), Y(V:\+\:) ]}}
> * Next step is removing replica {{A}}, so the result configuration is {{[
> B(V:\+\:), C(V:\+\:), Y(V:\+\:) ]}} instead of the expected {{[ B(V:\+\:),
> C(V:\+\:), X(V:\+\:) ]}}
> In this context, it's necessary to document that the 'target' replica is not
> the guaranteed destination of the replica move process, but just a pivot.
> Also, it make sense to update the CLI tool to accept only the source replica
> as an argument, where the target replica is selected by the system itself (if
> it's not so already).
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
