Attila Doroszlai created HDDS-8494:
--------------------------------------
Summary: Adjust replication queue limits for decommissioning nodes
Key: HDDS-8494
URL: https://issues.apache.org/jira/browse/HDDS-8494
Project: Apache Ozone
Issue Type: Sub-task
Components: Ozone Datanode, SCM
Reporter: Stephen O'Donnell
Assignee: Attila Doroszlai
When a node hosting a Ratis container is decommissioned, there are generally 3
sources available for the container replicas. One on the decommissioning host,
and then 2 others on somewhat random nodes across the cluster. This allows the
decommissioning load and hence speed of decommission to be shared across many
more nodes.
For an EC container, the decommissioning host is likely the only source of the
replica which needs to be copied and hence the decommission will be slower.
A host which is decommissioning is generally not used for Ratis reads unless
there are no other nodes available, but it would still be used for EC reads to
avoid online reconstruction. As decommission progresses on the node, and new
copies are formed, the read load will decline over time. Furthermore,
decommissioning nodes are not used for writes, so they should be under less
load than other cluster nodes.
Due to the reduced load on a decommissioning host, it is possible to increase
the number of commands queued on a decommissioning host and also increase the
size of the executor thread pool to process the commands.
When a datanode switches to a decommissioning state, it will adjust the size of
the replication supervisor thread pool higher, and if the node returns to the
In Service state, it will return to the lower thread pool limit.
Similarly when scheduling commands, SCM can allocate more commands to the
decommissioning host, as it should process them more quickly due to the lower
load and increased threadpool.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
