[ 
https://issues.apache.org/jira/browse/HDDS-17793?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Ivan Andika updated HDDS-17793:
-------------------------------
    Description: 
This is just an initial idea for discussion.

HDFS-7541 implements an upgrade domain to allow faster datanode rolling upgrade 
since it allows all the datanodes within an upgrade domain to be restarted 
altogether since the block placement policy (i.e. 
BlockPlacementPolicyWithUpgradeDomain) guarantee a block to have replicas 
across multiple upgrade domains. See 
[https://github.com/apache/hadoop/blob/trunk/hadoop-hdfs-project/hadoop-hdfs/src/site/markdown/HdfsUpgradeDomain.md]
 for the documentation.

Ozone currently does not guarantee this upgrade domain and AFAIK it is unsafe 
to batch restart datanodes (e.g. there might be containers with replicas in 
these batch).

Therefore, we are doing sequential restarts with this high level SOP:
 # Admin put datanode to ENTERING_MAINTENANCE
 # Admin wait until datanode is IN_MAINTENANCE since this signals that there 
are enough container replicas and won't affect read availability
 ## This can involves some container replication if putting the datanode to 
maintenance can cause data unavailability, etc
 # Admin restart the datanode
 # Admin recommission the datanode
 # Continue with the next datanode

I'm currently not sure whether Ozone supports restarting more than one 
datanodes without putting the datanode to maintenance and whether it's 
worthwhile to support Upgrade Domain feature or whether there are a SoTA 
implementation to handle this (with nice property like N datanodes can be 
restarted without manually specifying upgrade domain).

  was:
This is just an initial idea for discussion.

HDFS-7541 implements an upgrade domain to allow faster datanode rolling upgrade 
since it allows all the datanodes within a upgrade domain to be restarted 
altogether since the block placement policy (i.e. 
BlockPlacementPolicyWithUpgradeDomain). See 
[https://github.com/apache/hadoop/blob/trunk/hadoop-hdfs-project/hadoop-hdfs/src/site/markdown/HdfsUpgradeDomain.md]
 for the documentation.

Ozone currently does not guarantee this upgrade domain and AFAIK it is unsafe 
to batch restart datanodes.

Therefore, we are doing sequential restarts with this high level SOP:
 # Admin put datanode to ENTERING_MAINTENANCE
 # Admin wait until datanode is IN_MAINTENANCE since this signals that there 
are enough container replicas and won't affect read availability
 ## This can involves some container replication if putting the datanode to 
maintenance can cause data unavailability, etc
 # Admin restart the datanode
 # Admin recommission the datanode
 # Continue with the next datanode

I'm currently not sure whether Ozone supports restarting more than one 
datanodes without putting the datanode to maintenance and whether it's 
worthwhile to support Upgrade Domain feature or whether there are a SoTA 
implementation to handle this (with nice property like N datanodes can be 
restarted without manually specifying upgrade domain).


> Upgrade Domain in Ozone
> -----------------------
>
>                 Key: HDDS-17793
>                 URL: https://issues.apache.org/jira/browse/HDDS-17793
>             Project: Apache Ozone
>          Issue Type: New Feature
>            Reporter: Ivan Andika
>            Assignee: Ivan Andika
>            Priority: Major
>
> This is just an initial idea for discussion.
> HDFS-7541 implements an upgrade domain to allow faster datanode rolling 
> upgrade since it allows all the datanodes within an upgrade domain to be 
> restarted altogether since the block placement policy (i.e. 
> BlockPlacementPolicyWithUpgradeDomain) guarantee a block to have replicas 
> across multiple upgrade domains. See 
> [https://github.com/apache/hadoop/blob/trunk/hadoop-hdfs-project/hadoop-hdfs/src/site/markdown/HdfsUpgradeDomain.md]
>  for the documentation.
> Ozone currently does not guarantee this upgrade domain and AFAIK it is unsafe 
> to batch restart datanodes (e.g. there might be containers with replicas in 
> these batch).
> Therefore, we are doing sequential restarts with this high level SOP:
>  # Admin put datanode to ENTERING_MAINTENANCE
>  # Admin wait until datanode is IN_MAINTENANCE since this signals that there 
> are enough container replicas and won't affect read availability
>  ## This can involves some container replication if putting the datanode to 
> maintenance can cause data unavailability, etc
>  # Admin restart the datanode
>  # Admin recommission the datanode
>  # Continue with the next datanode
> I'm currently not sure whether Ozone supports restarting more than one 
> datanodes without putting the datanode to maintenance and whether it's 
> worthwhile to support Upgrade Domain feature or whether there are a SoTA 
> implementation to handle this (with nice property like N datanodes can be 
> restarted without manually specifying upgrade domain).



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to