[
https://issues.apache.org/jira/browse/HDFS-11701?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
James Moore updated HDFS-11701:
-------------------------------
Description:
We recently encountered the following NPE due to the DFSInputStream storing old
cached block locations from hosts which could no longer resolve.
{quote}
Caused by: java.lang.NullPointerException
at org.apache.hadoop.hdfs.DFSClient.isLocalAddress(DFSClient.java:1122)
at
org.apache.hadoop.hdfs.shortcircuit.DomainSocketFactory.getPathInfo(DomainSocketFactory.java:148)
at
org.apache.hadoop.hdfs.BlockReaderFactory.getBlockReaderLocal(BlockReaderFactory.java:474)
at
org.apache.hadoop.hdfs.BlockReaderFactory.build(BlockReaderFactory.java:354)
at
org.apache.hadoop.hdfs.DFSInputStream.blockSeekTo(DFSInputStream.java:662)
at
org.apache.hadoop.hdfs.DFSInputStream.seekToNewSource(DFSInputStream.java:1613)
at
org.apache.hadoop.fs.FSDataInputStream.seekToNewSource(FSDataInputStream.java:127)
~HBase related stack frames trimmed~
{quote}
After investigating, the DFSInputStream appears to have been open for upwards
of 3-4 weeks and had cached block locations from decommissioned nodes that no
longer resolve in DNS and had been shutdown and removed from the cluster 2
weeks prior. If the DFSInputStream had refreshed its block locations from the
name node, it would have received alternative block locations which would not
contain the decommissioned data nodes. As the above NPE leaves the
non-resolving data node in the list of block locations the DFSInputStream never
refreshes the block locations and all attempts to open a BlockReader for the
given blocks will fail.
In our case, we resolved the NPE by closing and re-opening every DFSInputStream
in the cluster to force a purge of the block locations cache. Ideally, the
DFSInputStream would re-fetch all block locations for a host which can't be
resolved in DNS or at least the blocks requested.
was:
We recently encountered the following NPE due to the DFSInputStream storing old
cached block locations from hosts which could no longer resolve.
```Caused by: java.lang.NullPointerException
at org.apache.hadoop.hdfs.DFSClient.isLocalAddress(DFSClient.java:1122)
at
org.apache.hadoop.hdfs.shortcircuit.DomainSocketFactory.getPathInfo(DomainSocketFactory.java:148)
at
org.apache.hadoop.hdfs.BlockReaderFactory.getBlockReaderLocal(BlockReaderFactory.java:474)
at
org.apache.hadoop.hdfs.BlockReaderFactory.build(BlockReaderFactory.java:354)
at
org.apache.hadoop.hdfs.DFSInputStream.blockSeekTo(DFSInputStream.java:662)
at
org.apache.hadoop.hdfs.DFSInputStream.seekToNewSource(DFSInputStream.java:1613)
at
org.apache.hadoop.fs.FSDataInputStream.seekToNewSource(FSDataInputStream.java:127)
~HBase related stack frames trimmed~```
After investigating, the DFSInputStream appears to have been open for upwards
of 3-4 weeks and had cached block locations from decommissioned nodes that no
longer resolve in DNS and had been shutdown and removed from the cluster 2
weeks prior. If the DFSInputStream had refreshed its block locations from the
name node, it would have received alternative block locations which would not
contain the decommissioned data nodes. As the above NPE leaves the
non-resolving data node in the list of block locations the DFSInputStream never
refreshes the block locations and all attempts to open a BlockReader for the
given blocks will fail.
In our case, we resolved the NPE by closing and re-opening every DFSInputStream
in the cluster to force a purge of the block locations cache. Ideally, the
DFSInputStream would re-fetch all block locations for a host which can't be
resolved in DNS or at least the blocks requested.
> NPE from Unresolved Host causes permanent DFSInputStream failures
> -----------------------------------------------------------------
>
> Key: HDFS-11701
> URL: https://issues.apache.org/jira/browse/HDFS-11701
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs-client
> Affects Versions: 2.7.0
> Environment: AWS Centos linux running HBase CDH 5.9.0 and HDFS CDH
> 5.9.0
> Reporter: James Moore
>
> We recently encountered the following NPE due to the DFSInputStream storing
> old cached block locations from hosts which could no longer resolve.
> {quote}
> Caused by: java.lang.NullPointerException
> at org.apache.hadoop.hdfs.DFSClient.isLocalAddress(DFSClient.java:1122)
> at
> org.apache.hadoop.hdfs.shortcircuit.DomainSocketFactory.getPathInfo(DomainSocketFactory.java:148)
> at
> org.apache.hadoop.hdfs.BlockReaderFactory.getBlockReaderLocal(BlockReaderFactory.java:474)
> at
> org.apache.hadoop.hdfs.BlockReaderFactory.build(BlockReaderFactory.java:354)
> at
> org.apache.hadoop.hdfs.DFSInputStream.blockSeekTo(DFSInputStream.java:662)
> at
> org.apache.hadoop.hdfs.DFSInputStream.seekToNewSource(DFSInputStream.java:1613)
> at
> org.apache.hadoop.fs.FSDataInputStream.seekToNewSource(FSDataInputStream.java:127)
> ~HBase related stack frames trimmed~
> {quote}
> After investigating, the DFSInputStream appears to have been open for upwards
> of 3-4 weeks and had cached block locations from decommissioned nodes that no
> longer resolve in DNS and had been shutdown and removed from the cluster 2
> weeks prior. If the DFSInputStream had refreshed its block locations from
> the name node, it would have received alternative block locations which would
> not contain the decommissioned data nodes. As the above NPE leaves the
> non-resolving data node in the list of block locations the DFSInputStream
> never refreshes the block locations and all attempts to open a BlockReader
> for the given blocks will fail.
> In our case, we resolved the NPE by closing and re-opening every
> DFSInputStream in the cluster to force a purge of the block locations cache.
> Ideally, the DFSInputStream would re-fetch all block locations for a host
> which can't be resolved in DNS or at least the blocks requested.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
