This is the documentation for Cloudera Manager 5.1.x. Documentation for other versions is available at Cloudera Documentation.

Known Issues and Workarounds in Cloudera Manager 5

The following sections describe the current known issues in Cloudera Manager 5.

NameNode Incorrectly Reports Missing Blocks During Rolling Upgrade

During a rolling upgrade to any of the CDH releases listed below, the NameNode may report missing blocks after rolling multiple DataNodes. This is caused by a race condition with block reporting between the DataNode and the NameNode. No permanent data loss occurs, but data can be unavailable for up to six hours before the problem corrects itself.

Releases affected: CDH 5.0.6, 5.1.5, 5.2.5, 5.3.3, 5.4.1, 5.4.2.

Releases containing the fix: CDH 5.2.6, 5.3.4, 5.4.3.

Workaround:
  • To avoid the problem - Cloudera advises skipping the affected releases and installing a release containing the fix. For example, do not upgrade to CDH 5.4.2; upgrade to CDH 5.4.3 instead.
  • If you have already completed an upgrade to an affected release, or are installing a new cluster - You can continue to run the release, or upgrade to a release that is not affected, as you choose. If you choose to upgrade to a release that is not affected, you must first upgrade to a version of Cloudera Manager—5.2.x, 5.3.x, or 5.4.x.—supporting the respective fixed CDH releases.

Using ext3 for server dirs easily hit inode limit

Using the ext3 filesystem for the Cloudera Manager command storage directory may exceed the maximum subdirectory size of 32000.

Workaround: Either decrease the value of the Command Eviction Age property so that the directories are more aggressively cleaned up, or migrate to the ext4 filesystem.

Hosts with Impala Llama roles must also have at least one YARN role

When integrated resource management is enabled for Impala, host(s) where the Impala Llama role(s) are running must have at least one YARN role. This is because Llama requires the topology.py script from the YARN configuration. If this requirement is not met, you may see errors such as:
"Exception running /etc/hadoop/conf.cloudera.yarn/topology.py
java.io.IOException: Cannot run program "/etc/hadoop/conf.cloudera.yarn/topology.py"
in the Llama role logs, and Impala queries may fail.

Workaround: Add a YARN gateway role to each Llama host that does not already have at least one YARN role (of any type).

The Spark History Server does not start when Kerberos authentication is enabled.

The Spark History Server does not start when managed by a Cloudera Manager 5.1 instance when Kerberos authentication is enabled.

Workaround:
  1. Go to the Spark service.
  2. Expand the History Server Default Group > Advanced category.
  3. Add the following configuration to the History Server Environment Advanced Configuration Snippet property:
    SPARK_HISTORY_OPTS=-Dspark.history.kerberos.enabled=true \
    -Dspark.history.kerberos.principal=principal \
    -Dspark.history.kerberos.keytab=keytab
where principal is the name of the Kerberos principal to use for the History Server, and keytab is the path to the principal's keytab file on the local filesystem of the host running the History Server.

Spurious warning on Accumulo 1.6 gateway hosts

When using the Accumulo shell on a host with only an Accumulo 1.6 Service gateway role, users will receive a warning about failing to create the directory /var/log/accumulo. The shell works normally otherwise.

Workaround: The warning is safe to ignore.

Accumulo 1.6 service log aggregation and search does not work

Cloudera Manager log aggregation and search features are incompatible with the log formatting needed by the Accumulo Monitor. Attempting to use either the "Log Search" diagnostics feature or the log file link off of an individual service role's summary page will result in empty search results.

Severity: High

Workaround: Operators can use the Accumulo Monitor to see recent severe log messages. They can see recent log messages below the WARNING level via a given role's process page and can inspect full logs on individual hosts by looking in /var/log/accumulo.

Cloudera Manager incorrectly sizes Accumulo Tablet Server max heap size after 1.4.4-cdh4.5.0 to 1.6.0-cdh4.6.0 upgrade

Because the upgrade path from Accumulo 1.4.4-cdh4.5.0 to 1.6.0-cdh4.6.0 involves having both services installed simultaneously, Cloudera Manager will be under the impression that worker hosts in the cluster are oversubscribed on memory and attempt to downsize the max heap size allowed for 1.6.0-cdh4.6.0 Tablet Servers.

Severity: High

Workaround: Manually verify that the Accumulo 1.6.0-cdh4.6.0 Tablet Server max heap size is large enough for your needs. Cloudera recommends you set this value to the sum of 1.4.4-cdh4.5.0 Tablet Server and Logger heap sizes.

Cluster CDH version not configured correctly for package installs

If you have installed CDH as a package, after the install make sure that the cluster CDH version matches the package CDH version. If the cluster CDH version does not match the package CDH version, Cloudera Manager will incorrectly enable and disable service features based on the cluster's configured CDH version.

Workaround:

Manually configure the cluster CDH version to match the package CDH version. Click ClusterName > Actions > Configure CDH Version. In the dialog, Cloudera Manager displays the installed CDH version, and asks for confirmation to configure itself with the new version.

Accumulo installations using LZO do not indicate dependence on the GPL Extras parcel

Accumulo 1.6 installations that use LZO compression functionality do not indicate that LZO depends on the GPL Extras parcel. When Accumulo is configured to use LZO, Cloudera Manager has no way to track that the Accumulo service now relies on the GPL Extras parcel. This prevents Cloudera Manager from warning administrators before they remove the parcel while Accumulo still requires it for proper operation.

Workaround: Check your Accumulo 1.6 service for the configuration changes mentioned in the Cloudera documentation for using Accumulo with CDH prior to removing the GPL Extras parcel. If the parcel is mistakenly removed, reinstall it and restart the Accumulo 1.6 service.

NodeManager Container Memory not imported to YARN correctly

When MapReduce configurations are imported to YARN, the property yarn.nodemanager.resource.memory-mb may not be set correctly for optimal YARN performance.

Workaround: Configure this property manually after running the import, then restart YARN.

Clients of the JobHistory Server Admin Interface Require Advanced Configuration Snippet

Clients of the JobHistory server administrative interface, such as the mapred hsadmin tool, may fail to connect to the server when run on hosts other than the one where the JobHistory server is running.

Workaround: Add the following to both the MapReduce Client Advanced Configuration Snippet for mapred-site.xml and the Cluster-wide Advanced Configuration Snippet for core-site.xml, replacing JOBHISTORY_SERVER_HOST with the hostname of your JobHistory server:
<property>
<name>mapreduce.history.admin.address</name>
<value>JOBHISTORY_SERVER_HOST:10033</value>
</property>

Created pools are not preserved when Dynamic Resource Pools page is used to configure YARN or Impala

Pools created on demand are not preserved when changes are made using the Dynamic Resource Pools page. If the Dynamic Resource Pools page is used to configure YARN and/or Impala services in a cluster, it is possible to specify pool placement rules that create a pool if one does not already exist. If changes are made to the configuration using this page, pools created as a result of such rules are not preserved across the configuration change.

Workaround: Submit the YARN application or Impala query as before, and the pool will be created on demand once again.

User should be prompted to add the AMON role when adding MapReduce to a CDH 5 cluster

When the MapReduce service is added to a CDH 5 cluster, the user is not asked to add the AMON role. Then, an error displays when the user tries to view MapReduce activities.

Workaround: Manually add the AMON role after adding the MapReduce service.

Enterprise license expiration alert not displayed until Cloudera Manager Server is restarted

When an enterprise license expires, the expiration notification banner is not displayed until the Cloudera Manager Server has been restarted (although the expired enterprise features stop working immediately upon license expiration, as expected).

Workaround: None.

Cluster installation with CDH 4.1 and Impala fails

In Cloudera Manager 5.0, installing a new cluster through the wizard with CDH 4.1 and Impala fails with the following error message, "dfs.client.use.legacy.blockreader.local is not enabled."

Workaround: Perform one of the following:
  1. Use CDH 4.2 or higher, or
  2. Install all desired services except Impala in your initial cluster setup. From the home page, use the drop-down menu near the cluster name and select Configure CDH Version. Confirm the version, then add Impala.

HDFS replication does not work from CDH 5 to CDH 4 with different realms

HDFS replication does not work from CDH 5 to CDH 4 with different realms. This is because authentication fails for services in a non-default realm via the WebHdfs API due to a JDK bug. This has been fixed in JDK6-u34 (b03)) and in JDK7.

Workaround: Use JDK 7 or upgrade JDK6 to at least version u34.

The Spark Upload Jar command fails in a secure cluster

The Spark Upload Jar command fails in a secure cluster.

Workaround: To run Spark on YARN, manually upload the Spark assembly jar to HDFS /user/spark/share/lib. The Spark assembly jar is located on the local filesystem, typically in /usr/lib/spark/assembly/lib or /opt/cloudera/parcels/CDH/lib/spark/assembly/lib.

Configurations for decommissioned roles not migrated from MapReduce to YARN

When the Import MapReduce Configuration wizard is used to import MapReduce configurations to YARN, decommissioned roles in the MapReduce service do not cause the corresponding imported roles to be marked as decommissioned in YARN.

Workaround: Delete or decommission the roles in YARN after running the import.

The HDFS command Roll Edits does not work in the UI when HDFS is federated

The HDFS command Roll Edits does not work in the Cloudera Manager UI when HDFS is federated because the command doesn't know which nameservice to use.

Workaround: Use the API, not the Cloudera Manager UI, to execute the Roll Edits command.

Cloudera Manager reports a confusing version number if you have oozie-client, but not oozie installed on a CDH 4.4 node

In CDH versions before 4.4, the metadata identifying Oozie was placed in the client, rather than the server package. Consequently, if the client package is not installed, but the server is, Cloudera Manager will report Oozie has been present but as coming from CDH 3 instead of CDH 4.

Workaround: Either install the oozie-client package, or upgrade to at least CDH 4.4. Parcel based installations are unaffected.

Cloudera Manager doesn't work with CDH 5.0.0 Beta 1

When you upgrade from Cloudera Manager 5.0.0 Beta 1 with CDH 5.0.0 Beta 1 to Cloudera Manager 5.0.0 Beta 2, Cloudera Manager won't work with CDH 5.0.0 Beta 1 and there's no notification of that fact.

Workaround: None. Do a new installation of CDH 5.0.0 Beta 2.

On CDH 4.1 secure clusters managed by Cloudera Manager 4.8.1 and higher, the Impala Catalog server needs advanced configuration snippet update

Impala queries fail on CDH 4.1 when Hive "Bypass Hive Metastore Server" option is selected.

Workaround: Add the following to Impala catalog server advanced configuration snippet for hive-site.xml, replacing Hive_Metastore_Server_Host with the host name of your Hive Metastore Server:

<property>
<name>hive.metastore.local</name>
<value>false</value>
</property>
<property>
<name>hive.metastore.uris</name>
<value>thrift://Hive_Metastore_Server_Host:9083</value>
</property>

Rolling Upgrade to CDH 5 is not supported.

Rolling upgrade between CDH 4 and CDH 5 is not supported. Incompatibilities between major versions means rolling restarts are not possible. In addition, rolling upgrade will not be supported from CDH 5.0.0 Beta 1 to any later releases, and may not be supported between any future beta versions of CDH 5 and the General Availability release of CDH 5.

Workaround: None.

Error reading .zip file created with the Collect Diagnostic Data command.

After collecting Diagnostic Data and using the Download Diagnostic Data button to download the created zip file to the local system, the zip file cannot be opened using the FireFox browser on a Macintosh. This is because the zip file is created as a Zip64 file, and the unzip utility included with Macs does not support Zip64. The zip utility must be version 6.0 or later. You can determine the zip version with unzip -v.

Workaround: Update the unzip utility to a version that supports Zip64.

Enabling wildcarding in a secure environment causes NameNode to fail to start.

In a secure cluster, you cannot use a wildcard for the NameNode's RPC or HTTP bind address, or the NameNode will fail to start. For example, dfs.namenode.http-address must be a real, routable address and port, not 0.0.0.0.port. In Cloudera Manager, the "Bind NameNode to Wildcard Address" property must not be enabled. This should affect you only if you are running a secure cluster and your NameNode needs to bind to multiple local addresses.

Bug: HDFS-4448

Severity: Medium

Workaround: Disable the "Bind NameNode to Wildcard Address" property found under the Configuration tab for the NameNode role group.

After JobTracker failover, complete jobs from the previous active JobTracker are not visible.

When a JobTracker failover occurs and a new JobTracker becomes active, the new JobTracker UI does not show the completed jobs from the previously active JobTracker (that is now the standby JobTracker). For these jobs the "Job Details" link does not work.

Severity: Med

Workaround: None.

After JobTracker failover, information about rerun jobs is not updated in Activity Monitor.

When a JobTracker failover occurs while there are running jobs, jobs are restarted by the new active JobTracker by default. For the restarted jobs the Activity Monitor will not update the following: 1) The start time of the restarted job will remain the start time of the original job. 2) Any Map or Reduce task that had finished before the failure happened will not be updated with information about the corresponding task that was rerun by the new active JobTracker.

Severity: Med

Workaround: None.

Installing on AWS, you must use private EC2 hostnames.

When installing on an AWS instance, and adding hosts using their public names, the installation will fail when the hosts fail to heartbeat.

Severity: Med

Workaround:

Use the Back button in the wizard to return to the original screen, where it prompts for a license.

Rerun the wizard, but choose "Use existing hosts" instead of searching for hosts. Now those hosts show up with their internal EC2 names.

Continue through the wizard and the installation should succeed.

After removing and then re-adding a service, the alternatives settings are incorrect.

After deleting a service, the alternatives settings are not cleaned up. If you then re-add the service, it will be given a new instance name, and a new set of configurations settings are added. However, because both the new and old (deleted) instances have the same alternatives priority, the original one will be used rather than the newer one.

Severity: Med

Workaround: The simplest way to fix this is:

  1. Go to the Configuration tab for the new service instance in Cloudera Manager
  2. Search for "alternatives".
  3. Raise the priority value and save the setting.
  4. Redeploy the client configuration.

If HDFS uses Quorum-based Storage without HA enabled, the SecondaryNameNode cannot checkpoint.

If HDFS is set up in non-HA mode, but with Quorum-based storage configured, the dfs.namenode.edits.dir is automatically configured to the Quorum-based Storage URI. However, the SecondaryNameNode cannot currently read the edits from a Quorum-based Storage URI, and will be unable to do a checkpoint.

Severity: Medium

Workaround: Add to the NameNode's advanced configuration snippet the dfs.namenode.edits.dir property with both the value of the Quorum-based Storage URI as well as a local directory, and restart the NameNode. For example,

<property> <name>dfs.namenode.edits.dir</name>
<value>qjournal://jn1HostName:8485;jn2HostName:8485;jn3HostName:8485/journalhdfs1,file:///dfs/edits</value>
</property>

Changing the rack configuration may temporarily cause mis-replicated blocks to be reported.

A rack re-configuration will cause HDFS to report mis-replicated blocks until HDFS rebalances the system, which may take some time. This is a normal side-effect of changing the configuration.

Severity: Low

Workaround: None

Cannot use '/' as a mount point with a Federated HDFS Nameservice.

A Federated HDFS Service doesn't support nested mount points, so it is impossible to mount anything at '/'. Because of this issue, the root directory will always be read-only, and any client application that requires a writeable root directory will fail.

Severity: Low

Workaround:
  1. In the CDH 4 HDFS Service > Configuration tab of the Cloudera Manager Admin Console, search for "nameservice".
  2. In the Mountpoints field, change the mount point from "/" to a list of mount points that are in the namespace that the Nameservice will manage. (You can enter this as a comma-separated list - for example, "/hbase, /tmp, /user" or by clicking the plus icon to add each mount point in its own field.) You can determine the list of mount points by running the command hadoop fs -ls / from the CLI on the NameNode host.

Historical disk usage reports do not work with federated HDFS.

Severity: Low

Workaround: None.

(CDH 4 only) Activity monitoring does not work on YARN activities.

Activity monitoring is not supported for YARN in CDH 4.

Severity: Low

Workaround: None

HDFS monitoring configuration applies to all Nameservices

The monitoring configurations at the HDFS level apply to all Nameservices. So, if there are two federated Nameservices, it's not possible to disable a check on one but not the other. Likewise, it's not possible to have different thresholds for the two Nameservices.

Severity: Low

Workaround: None

Supported and Unsupported Replication Scenarios and Limitations

Restoring snapshot of a file to an empty directory does not overwrite the directory

Restoring the snapshot of an HDFS file to an HDFS path that is an empty HDFS directory (using the Restore As action) will result in the restored file present inside the HDFS directory instead of overwriting the empty HDFS directory.

Workaround: None.

HDFS Snapshot appears to fail if policy specifies duplicate directories.

In an HDFS snapshot policy, if a directory is specified more than once, the snapshot appears to fail with an error message on the Snapshot page. However, in the HDFS Browser, the snapshot is shown as having been created successfully.

Severity: Low

Workaround: Remove the duplicate directory specification from the policy.

Hive replication fails if "Force Overwrite" is not set.

The Force Overwrite option, if checked, forces overwriting data in the target metastore if there are incompatible changes detected. For example, if the target metastore was modified and a new partition was added to a table, this option would force deletion of that partition, overwriting the table with the version found on the source. If the Force Overwrite option is not set, recurring replications may fail.

Severity: Med

Workaround: Set the Force Overwrite option.

During HDFS replication, tasks may fail due to DataNode timeouts.

In CDH 4.2, during an HDFS replication job (using Cloudera Manager's Backup and Data Recovery product) individual tasks in the Replication job may fail due to DataNode timeouts. If enough of these timeouts occur, the replication task may slow down, and the entire replication job could time out and fail.

Severity: Medium

Workaround: None.

Page generated September 3, 2015.