hfutatzhanghb commented on PR #5889: URL: https://github.com/apache/hadoop/pull/5889#issuecomment-1651135519
@Hexiaoqiao @zhangshuyan0 Sir, thanks for replying. Yes, I met some issues. Details are below: 1. As mentioned in [HDFS-17047](https://issues.apache.org/jira/browse/HDFS-17047), we found two ec blocks of one same block group are in the different storages of one datanode. This case can cause ec reconstruction infinitely. 2. To resolve this problem, We decide to use DirectoryScanner to detect duplicated ec internal blocks and delete them. But delete them directly in any order is not safe, because the datanode generates fbr using unordered storages index. We can not infer which duplicated replica will be handled by `BlockManager#addStoredBlock` method lastly, So I think we should sort datanode storages when generating fbr to make sure we can use this order in `FsDatasetImpl#checkAndUpdate` method to delete the replicas which handled by namenode earlier. 3. And when deleteing duplicated replicas, We only delete those scanInfos whose (now - time be scanned) > block report interval to make sure we do delete safely. Hope to receive your suggestions~ Thanks a lot. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
