[
https://issues.apache.org/jira/browse/HDFS-11398?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Yiqun Lin updated HDFS-11398:
-----------------------------
Description:
The test {{TestDataNodeVolumeFailure#testUnderReplicationAfterVolFailure}}
still fails intermittently in trunk after HDFS-11316. The stack infos:
{code}
java.util.concurrent.TimeoutException: Timed out waiting for DN to die
at
org.apache.hadoop.hdfs.DFSTestUtil.waitForDatanodeDeath(DFSTestUtil.java:702)
at
org.apache.hadoop.hdfs.server.datanode.TestDataNodeVolumeFailureReporting.testSuccessiveVolumeFailures(TestDataNodeVolumeFailureReporting.java:218
{code}
I looked into this and found there is one chance that the vaule
{{UnderReplicatedBlocksCount}} will be no longer > 0. The following is my
analysation:
In test {{TestDataNodeVolumeFailure.testUnderReplicationAfterVolFailure}}, it
uses creating file to trigger the disk error checking. The related codes:
{code}
Path file1 = new Path("/test1");
DFSTestUtil.createFile(fs, file1, 1024, (short)3, 1L);
DFSTestUtil.waitReplication(fs, file1, (short)3);
// Fail the first volume on both datanodes
File dn1Vol1 = new File(dataDir, "data"+(2*0+1));
File dn2Vol1 = new File(dataDir, "data"+(2*1+1));
DataNodeTestUtils.injectDataDirFailure(dn1Vol1, dn2Vol1);
Path file2 = new Path("/test2");
DFSTestUtil.createFile(fs, file2, 1024, (short)3, 1L);
DFSTestUtil.waitReplication(fs, file2, (short)3);
{code}
This will lead one problem: If the cluster is busy, and it costs long time to
wait replication of file2 to be desired value. During this time, the under
replication blocks of file1 can also be rereplication in cluster. If this is
done, the condition {{underReplicatedBlocks > 0}} will never be satisfied.
And this can be reproduced in my local env.
Actually, we can use a easy way {{DataNodeTestUtils.waitForDiskError}} to
replace this, it runs fast and be more reliable.
was:
The test {{TestDataNodeVolumeFailure#testUnderReplicationAfterVolFailure}}
still fails intermittently in trunk after HDFS-11316. The stack infos:
{code}
java.util.concurrent.TimeoutException: Timed out waiting for DN to die
at
org.apache.hadoop.hdfs.DFSTestUtil.waitForDatanodeDeath(DFSTestUtil.java:702)
at
org.apache.hadoop.hdfs.server.datanode.TestDataNodeVolumeFailureReporting.testSuccessiveVolumeFailures(TestDataNodeVolumeFailureReporting.java:218
{code}
I looked into this and found there is one chance that the vaule
{{UnderReplicatedBlocksCount}} will be no longer > 0. The following is my
analysation:
In test {{TestDataNodeVolumeFailure.testUnderReplicationAfterVolFailure}}, it
uses creating file to trigger the disk error checking. The related codes:
{code}
Path file1 = new Path("/test1");
DFSTestUtil.createFile(fs, file1, 1024, (short)3, 1L);
DFSTestUtil.waitReplication(fs, file1, (short)3);
// Fail the first volume on both datanodes
File dn1Vol1 = new File(dataDir, "data"+(2*0+1));
File dn2Vol1 = new File(dataDir, "data"+(2*1+1));
DataNodeTestUtils.injectDataDirFailure(dn1Vol1, dn2Vol1);
Path file2 = new Path("/test2");
DFSTestUtil.createFile(fs, file2, 1024, (short)3, 1L);
DFSTestUtil.waitReplication(fs, file2, (short)3);
{code}
This will lead one problem: If the cluster is busy, and it costs long time to
wait replication of file2 to be desired value. And there is one chance that the
under replication blocks of file1 can also be rereplication in cluster. If this
is done, the condition {{underReplicatedBlocks > 0}} will never be satisfied.
And this can be reproduced in my local env.
Actually, we can use a easy way {{DataNodeTestUtils.waitForDiskError}} to
replace this, it runs fast and be more reliable.
> TestDataNodeVolumeFailure#testUnderReplicationAfterVolFailure still fails
> intermittently
> ----------------------------------------------------------------------------------------
>
> Key: HDFS-11398
> URL: https://issues.apache.org/jira/browse/HDFS-11398
> Project: Hadoop HDFS
> Issue Type: Bug
> Affects Versions: 3.0.0-alpha2
> Reporter: Yiqun Lin
> Assignee: Yiqun Lin
> Attachments: failure.log, HDFS-11398.001.patch,
> HDFS-11398-reproduce.patch
>
>
> The test {{TestDataNodeVolumeFailure#testUnderReplicationAfterVolFailure}}
> still fails intermittently in trunk after HDFS-11316. The stack infos:
> {code}
> java.util.concurrent.TimeoutException: Timed out waiting for DN to die
> at
> org.apache.hadoop.hdfs.DFSTestUtil.waitForDatanodeDeath(DFSTestUtil.java:702)
> at
> org.apache.hadoop.hdfs.server.datanode.TestDataNodeVolumeFailureReporting.testSuccessiveVolumeFailures(TestDataNodeVolumeFailureReporting.java:218
> {code}
> I looked into this and found there is one chance that the vaule
> {{UnderReplicatedBlocksCount}} will be no longer > 0. The following is my
> analysation:
> In test {{TestDataNodeVolumeFailure.testUnderReplicationAfterVolFailure}}, it
> uses creating file to trigger the disk error checking. The related codes:
> {code}
> Path file1 = new Path("/test1");
> DFSTestUtil.createFile(fs, file1, 1024, (short)3, 1L);
> DFSTestUtil.waitReplication(fs, file1, (short)3);
> // Fail the first volume on both datanodes
> File dn1Vol1 = new File(dataDir, "data"+(2*0+1));
> File dn2Vol1 = new File(dataDir, "data"+(2*1+1));
> DataNodeTestUtils.injectDataDirFailure(dn1Vol1, dn2Vol1);
> Path file2 = new Path("/test2");
> DFSTestUtil.createFile(fs, file2, 1024, (short)3, 1L);
> DFSTestUtil.waitReplication(fs, file2, (short)3);
> {code}
> This will lead one problem: If the cluster is busy, and it costs long time to
> wait replication of file2 to be desired value. During this time, the under
> replication blocks of file1 can also be rereplication in cluster. If this is
> done, the condition {{underReplicatedBlocks > 0}} will never be satisfied.
> And this can be reproduced in my local env.
> Actually, we can use a easy way {{DataNodeTestUtils.waitForDiskError}} to
> replace this, it runs fast and be more reliable.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
