[
https://issues.apache.org/jira/browse/IGNITE-24383?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vladimir Pligin updated IGNITE-24383:
-------------------------------------
Description:
h3. Motivation
Currently primary replica is not capable of being gracefully transferred while
assignments remain stable. This could imply uneven load in a cluster for long
periods of time - for example when a user upscales the cluster, in a situation
like that replicas won't be spread across the entire set of nodes, the replicas
will remain on the pre-scaled set of nodes.
Considerations
* Implementing a graceful transfer (including locks) is a big feature, may be
practical, but needs additional thoughts.
* Automatic forcible transfer would imply unavailability measured in seconds
(up to lease interval). It's unacceptable in terms of UX.
* We may consider implementing an endpoint to manually reassign replicas. It
will give users a way to fix the uneven load.
h3. Definition of Done
* There's a management API to transfer all replicas at once.
> Add forcible primary replica transfer
> --------------------------------------
>
> Key: IGNITE-24383
> URL: https://issues.apache.org/jira/browse/IGNITE-24383
> Project: Ignite
> Issue Type: Improvement
> Reporter: Vladimir Pligin
> Priority: Major
> Labels: ignite-3
>
> h3. Motivation
> Currently primary replica is not capable of being gracefully transferred
> while assignments remain stable. This could imply uneven load in a cluster
> for long periods of time - for example when a user upscales the cluster, in a
> situation like that replicas won't be spread across the entire set of nodes,
> the replicas will remain on the pre-scaled set of nodes.
> Considerations
> * Implementing a graceful transfer (including locks) is a big feature, may
> be practical, but needs additional thoughts.
> * Automatic forcible transfer would imply unavailability measured in seconds
> (up to lease interval). It's unacceptable in terms of UX.
> * We may consider implementing an endpoint to manually reassign replicas. It
> will give users a way to fix the uneven load.
> h3. Definition of Done
> * There's a management API to transfer all replicas at once.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
