[ https://issues.apache.org/jira/browse/HDFS-12354?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16143402#comment-16143402 ]
Yiqun Lin commented on HDFS-12354: ---------------------------------- Thanks for the comments, [~cheersyang]. bq. Do you think if we can extend ContainerStatus to have another field such as numOfPendingDeletionBlocks? Adding new field in ContainerStatus seems need some more adjustment work for this. I noticed one detail that you mentioned there might more than 4000 containers in a datanode. Like this scenery, we will still meets some problems if we support get a shuffle list. Each time we only chose specified number of containers, there is a chance that some containers will not be chosen in a long time when the specified number is small while the whole containers is large. So based on this, I am thinking can we use the round-robin way? We can pass previous key and invoke {{containerMap#tailMap}} to select the next containers that these container name is great than previous key. For the pending under blocks based way, I prefer to implement this later and firstly implement above two ways (The randomly choosing way and round-robin way). These two ways only need minor change but will have a good effect on container choosing. > Ozone: Shuffle container list for datanode BlockDeletingService > --------------------------------------------------------------- > > Key: HDFS-12354 > URL: https://issues.apache.org/jira/browse/HDFS-12354 > Project: Hadoop HDFS > Issue Type: Sub-task > Components: datanode, ozone > Reporter: Weiwei Yang > Assignee: Yiqun Lin > > {{BlockDeletingService}} is a per-datanode container block deleting service > takes in charge of the "real" deletion of ozone blocks. It spawns a worker > thread per container and delete blocks/chunks from disk as background > threads. The number of threads currently is throttled by > {{ozone.block.deleting.container.limit.per.interval}}, but there is a > potential problem. Containers are sorted so it always fetch same of > containers, we need to fix this by creating an API in > {{ContainerManagerImpl}} to get a shuffled list of containers. -- This message was sent by Atlassian JIRA (v6.4.14#64029) --------------------------------------------------------------------- To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org