[
https://issues.apache.org/jira/browse/HDDS-15014?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Wei-Chiu Chuang updated HDDS-15014:
-----------------------------------
Release Note:
Sped up EC DataNode decommission by allowing SCM to switch overloaded
decommission workloads from 1:1 replication off the leaving node to
multi-source EC reconstruction using replicas on other DataNodes. DataNodes may
use per-disk push-replication thread pools so one slow disk does not block the
others.
*New configuration* (defaults preserve prior behavior):
* {{hdds.scm.replication.decommission.ec.reconstruction.enabled}} (default
{{false}}) — set {{true}} on SCM to enable;
{{hdds.scm.replication.decommission.ec.reconstruction.load.factor}} ({{0.9}});
{{hdds.scm.replication.reconstruction.global.limit}} ({{0}} = no cluster cap).
SCM keys are reconfigurable.
* {{hdds.datanode.replication.per.volume.enabled}} (default {{false}});
{{hdds.datanode.replication.per.volume.streams.limit}} ({{2}}).
No CLI or REST API changes. New SCM replication-manager metrics cover
reconstruction during decommission (e.g.
{{ecReconstructionDecommissionTriggeredTotal}}).
> Speed up EC container decommission
> ----------------------------------
>
> Key: HDDS-15014
> URL: https://issues.apache.org/jira/browse/HDDS-15014
> Project: Apache Ozone
> Issue Type: Epic
> Components: SCM
> Reporter: Wei-Chiu Chuang
> Assignee: Wei-Chiu Chuang
> Priority: Major
> Labels: pull-request-available
> Fix For: 2.3.0
>
>
> When Ozone decommission a datanode, it must wait until all containers are
> fully replicated to other datanodes in the cluster.
>
> In the case of EC containers, instead of reconstructing the block, it
> replicates to a destination. That replication becomes a bottleneck because it
> doesn't matter how big the cluster is, the container block is always
> replicated from the same source (the decommissioning datanode)
>
> Solution: keep doing replication, but when there are too many containers
> pending, do reconstruction instead. Reconstruction leverages other datanodes
> in the cluster so it increases parallelism.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]