Known Issues in HDFS

Learn about the known issues in HDFS, the impact or changes to the functionality, and the workaround.

OPSAPS-60958: The dfs.access.time.precision and dfs.namenode.accesstime.precision parameters are available in Cloudera Manager > HDFS > Configuration.
You must configure both the dfs.access.time.precision and dfs.namenode.accesstime.precision parameters with the same value as Cloudera Manager still sends both the parameters to HDFS service configuration.
ENGESC-19334: After configuring multiple NameNodes, with heavy read and write workloads in the cluster, there are chances of performance impact in terms of slowness on the client side.
This is because of the additional NN retry or probe introduced by the extra NN.
There are two solutions to mitigate the performance impact:
  • By reducing the value of the property: dfs.client.failover.sleep.base.millis // default is 500ms.
  • By increasing the value of the property dfs.client.failover.max.attempts // default is 15.
  • Using org.apache.hadoop.hdfs.server.namenode.ha.RequestHedgingProxyProvider instead of org.apache.hadoop.hdfs.server.namenode.ha.ConfiguredFailoverProxyProvider in HDFS client failover proxy provider (defined in dfs.client.failover.proxy.provider).
OPSAPS-64307: In the case when on a cluster the JournalNodes were restarted recently, the "Add new NameNode" wizard for HDFS service might fail to bootstrap the new NameNode, if there wasn't a new fsImage created since the restart of the JournalNodes, but during restarting them the edit logs were rolled in the system.
If the bootstrap fails during the "Add new NameNode" wizard, then do the following steps:
  1. Delete the newly added NameNode and FailoverController
  2. Move the active HDFS NameNode to safe mode
  3. Do a Save Namespace operation on the active HDFS NameNode
  4. Leave safe mode on the active HDFS NameNode
  5. Try to add the new NameNode again
OPSAPS-64363: Deleting of additional Standby Namenode does not delete the ZKFC role and this has to be done manually.
OPSAPS-63558: Snapshot diff based HDFS replications do not provide correct file delete and rename counters through the API.
The number of files deleted and renamed by DistCp for snapshot based replications can be checked in the logs provided by DistCp on the standard error output.
CDPD-28459: After performing an upgrade rollback from CDP 7.1.7 to CDH6, you may see the following error when restarting the DataNodes: ERROR datanode.DataNode: Exception in secureMain The path component: '/var/run/hdfs-sockets' in '/var/run/hdfs-sockets/dn' has permissions 0755 uid 39998 and gid 1006. It is not protected because it is owned by a user who is not root and not the effective user: '0'.
You must run the command described in the error message "chown root /var/run/hdfs-sockets". After this, the DataNode will restart successfully.
CDPD-28390: Rolling restart of the HDFS JournalNodes may time out on Ubuntu20.
If the restart operation times out, you can manually stop and restart the Name Node and Journal Node services one by one.
OPSAPS-60832: When decommission of DN runs for a longer time and when decommission monitor's kerberos ticket expires, it is not auto-renewed. Decommission of DN is not completed in CM as decommission monitor fails to fetch the state of DN after kerberos ticket expiry.
Decommission state of DN can be fetched using CLI command, i.e, hdfs dfsadmin -report.
OPSAPS-55788: WebHDFS is always enabled. The Enable WebHDFS checkbox does not take effect.
OPSAPS-63299: Disable HA command for a nameservice does not work if the nameservice has more than 2 NNs defined.
OPSAPS-63301: Deleting nameservice command does not delete all the NNs belonging to the nameservice, if there are more than two NNs that are assigned to the nameservice.
CDPD-50044: Data node tab loading issue in the name node UI
When clicking on the data node tab, the message NameNode is still loading. Redirecting to the Startup Progress page appears.
Unsupported Features
The following HDFS features are currently not supported in Cloudera Data Platform:

Technical Service Bulletins

TSB 2023-666: Out of order HDFS snapshot deletion may delete renamed/moved files, which may result in data loss
Cloudera has discovered a bug in the Apache Hadoop Distributed File System (HDFS) snapshot implementation. Deleting an HDFS snapshot may incorrectly remove files in the .Trash directories or remove renamed files from the current file system state. This is an unexpected behavior because deleting an HDFS snapshot should only delete the files stored in the specified snapshot, but not data in the current state.

In the particular HDFS installation in which the bug was discovered, deleting one of the snapshots caused certain files to be moved to trash and deletion of some of the files in a .Trash directory. Although it is clear that the conditions of the bug are (1) out-of-order snapshot deletion and (2) files moved to trash or other directories, we were unable to replicate the bug in other HDFS installations after executing similar test operations with a variety of different sequences. We also did not observe any actual data loss in our tests. However, there is a remote possibility that this bug may lead to data loss.

Knowledge article
For the latest update on this issue see the corresponding Knowledge article: TSB 2023-666: Out of order HDFS snapshot deletion may delete renamed/moved files, which may result in data loss
TSB 2022-549: Possible HDFS Erasure Coded (EC) data loss when EC blocks are over-replicated
Cloudera has detected a bug that can cause loss of data that is stored in HDFS Erasure Coded (EC) files in an unlikely scenario.

Some EC blocks may be inadvertently deleted due to a bug in how the NameNode chooses excess or over-replicated block replicas for deletion. One possible cause of over-replication is running the HDFS balancer soon after a NameNode goes into failover mode.

In a rare situation, the redundant blocks could be placed in such a way that one replica is in one rack, and few redundant replicas are in the same rack. Such placement causes a counting bug (HDFS-16420) to be triggered. Instead of deleting just the redundant replicas, the original replica may also be deleted.

Usually this is not an issue, because the lost replica can be detected and reconstructed from the remaining data and parity blocks. However, if multiple blocks in an EC Block Group are affected by this counting bug within a short time, the block cannot be reconstructed anymore. For example, 4 blocks are affected out of 9 for the RS(6,3) policy.

Another situation is recommissioning multiple nodes back into the same rack of the cluster where the current live replica exists.

Upstream JIRA
Knowledge article
For the latest update on this issue see the corresponding Knowledge article: TSB 2022-549: Possible HDFS Erasure Coded (EC) data loss when EC blocks are over-replicated