[jira] [Updated] (HDFS-7065) Pipeline close recovery race can cause block corruption
[ https://issues.apache.org/jira/browse/HDFS-7065?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Kihwal Lee updated HDFS-7065: - Resolution: Fixed Fix Version/s: 2.6.0 Status: Resolved (was: Patch Available) Thanks for the review, Nicholas. I've committed this to trunk and branch-2. Pipeline close recovery race can cause block corruption --- Key: HDFS-7065 URL: https://issues.apache.org/jira/browse/HDFS-7065 Project: Hadoop HDFS Issue Type: Bug Components: datanode Affects Versions: 2.5.0 Reporter: Kihwal Lee Assignee: Kihwal Lee Priority: Critical Fix For: 2.6.0 Attachments: HDFS-7065.patch If multiple pipeline close recoveries are performed against the same block, the replica may go corrupt. Here is one case I have observed: The client tried to close a block, but the ACK timed out. It excluded the first DN and tried pipeline recovery (recoverClose). It too failed and another recovery was attempted with only one DN. This took more than usual but the client eventually got an ACK and the file was closed successfully. Later on the one and only replica was found to be corrupt. It turned out the DN was having transient slow disk I/O issue at that time. The first recovery was stuck until the second recovery was attempted 30 seconds later. After few seconds, they both threads started running. The second recovery finished first and then the first recovery with an older gen stamp finished, turning gen stamp backward. There is a sanity check in {{recoverCheck()}}, but since check and modify are not synchronized, {{recoverClose()}} is not multi-thread safe. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Updated] (HDFS-7065) Pipeline close recovery race can cause block corruption
[ https://issues.apache.org/jira/browse/HDFS-7065?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Tsz Wo Nicholas Sze updated HDFS-7065: -- Component/s: datanode Hadoop Flags: Reviewed Hi Kihwal, thanks for working this. Are you able to reproduce the problem? I think this is hard to test. Anyway, you analysis makes sense and the patch looks good. +1 Pipeline close recovery race can cause block corruption --- Key: HDFS-7065 URL: https://issues.apache.org/jira/browse/HDFS-7065 Project: Hadoop HDFS Issue Type: Bug Components: datanode Affects Versions: 2.5.0 Reporter: Kihwal Lee Assignee: Kihwal Lee Priority: Critical Attachments: HDFS-7065.patch If multiple pipeline close recoveries are performed against the same block, the replica may go corrupt. Here is one case I have observed: The client tried to close a block, but the ACK timed out. It excluded the first DN and tried pipeline recovery (recoverClose). It too failed and another recovery was attempted with only one DN. This took more than usual but the client eventually got an ACK and the file was closed successfully. Later on the one and only replica was found to be corrupt. It turned out the DN was having transient slow disk I/O issue at that time. The first recovery was stuck until the second recovery was attempted 30 seconds later. After few seconds, they both threads started running. The second recovery finished first and then the first recovery with an older gen stamp finished, turning gen stamp backward. There is a sanity check in {{recoverCheck()}}, but since check and modify are not synchronized, {{recoverClose()}} is not multi-thread safe. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Updated] (HDFS-7065) Pipeline close recovery race can cause block corruption
[ https://issues.apache.org/jira/browse/HDFS-7065?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Kihwal Lee updated HDFS-7065: - Attachment: HDFS-7065.patch Pipeline close recovery race can cause block corruption --- Key: HDFS-7065 URL: https://issues.apache.org/jira/browse/HDFS-7065 Project: Hadoop HDFS Issue Type: Bug Affects Versions: 2.5.0 Reporter: Kihwal Lee Priority: Critical Attachments: HDFS-7065.patch If multiple pipeline close recoveries are performed against the same block, the replica may go corrupt. Here is one case I have observed: The client tried to close a block, but the ACK timed out. It excluded the first DN and tried pipeline recovery (recoverClose). It too failed and another recovery was attempted with only one DN. This took more than usual but the client eventually got an ACK and the file was closed successfully. Later on the one and only replica was found to be corrupt. It turned out the DN was having transient slow disk I/O issue at that time. The first recovery was stuck until the second recovery was attempted 30 seconds later. After few seconds, they both threads started running. The second recovery finished first and then the first recovery with an older gen stamp finished, turning gen stamp backward. There is a sanity check in {{recoverCheck()}}, but since check and modify are not synchronized, {{recoverClose()}} is not multi-thread safe. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Updated] (HDFS-7065) Pipeline close recovery race can cause block corruption
[ https://issues.apache.org/jira/browse/HDFS-7065?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Kihwal Lee updated HDFS-7065: - Status: Patch Available (was: Open) Pipeline close recovery race can cause block corruption --- Key: HDFS-7065 URL: https://issues.apache.org/jira/browse/HDFS-7065 Project: Hadoop HDFS Issue Type: Bug Affects Versions: 2.5.0 Reporter: Kihwal Lee Priority: Critical Attachments: HDFS-7065.patch If multiple pipeline close recoveries are performed against the same block, the replica may go corrupt. Here is one case I have observed: The client tried to close a block, but the ACK timed out. It excluded the first DN and tried pipeline recovery (recoverClose). It too failed and another recovery was attempted with only one DN. This took more than usual but the client eventually got an ACK and the file was closed successfully. Later on the one and only replica was found to be corrupt. It turned out the DN was having transient slow disk I/O issue at that time. The first recovery was stuck until the second recovery was attempted 30 seconds later. After few seconds, they both threads started running. The second recovery finished first and then the first recovery with an older gen stamp finished, turning gen stamp backward. There is a sanity check in {{recoverCheck()}}, but since check and modify are not synchronized, {{recoverClose()}} is not multi-thread safe. -- This message was sent by Atlassian JIRA (v6.3.4#6332)