[
https://issues.apache.org/jira/browse/HDFS-17156?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17753235#comment-17753235
]
Takanobu Asanuma commented on HDFS-17156:
-----------------------------------------
It looked like a bug in RBF SBN, so I moved this jira from HADOOP to HDFS.
> mapreduce job encounters java.io.IOException when
> dfs.client.rbf.observer.read.enable is true
> ---------------------------------------------------------------------------------------------
>
> Key: HDFS-17156
> URL: https://issues.apache.org/jira/browse/HDFS-17156
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: rbf
> Reporter: Chunyi Yang
> Priority: Minor
> Labels: Observer, RBF
>
> While executing a mapreduce job in an environment utilizing Router-Based
> Federation with Observer read enabled, there is an estimated 1% chance of
> encountering the following error.
> {code:java}
> "java.io.IOException: Resource
> hdfs://XXXX/user/XXXX/.staging/job_XXXXXX/.tez/application_XXXXXX/tez-conf.pb
> changed on src filesystem - expected: \"2023-07-07T12:41:16.801+0900\", was:
> \"2023-07-07T12:41:16.822+0900\", current time:
> \"2023-07-07T12:41:22.386+0900\"",
> {code}
> This error happens in function verifyAndCopy inside FSDownload.java when
> nodemanager tries to download a file right after the file has been written to
> the HDFS. The write operation runs on active namenode and read operation runs
> on observer namenode as expected.
> The edits file and hdfs-audit files indicate that the expected timestamp
> mentioned in the error message aligns with the OP_CLOSE MTIME of the
> 'tez-conf.pb' file (which is correct). However, the actual timestamp
> retrieved from the read operation corresponds to the OP_ADD MTIME of the
> target 'tez-conf.pf' file (which is incorrect). This inconsistency suggests
> that the observer namenode responds to the client before its edits file is
> updated with the latest stateId.
> Further troubleshooting has revealed that during write operations, the router
> responds to the client before receiving the latest stateId from the active
> namenode. Consequently, the outdated stateId is then used in the subsequent
> read operation on the observer namenode, leading to inaccuracies in the
> information provided by the observer namenode.
> To resolve this issue, it is essential to ensure that the router sends a
> response to the client only after receiving the latest stateId from the
> active namenode.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
