Known Issues and Limitations in CDH 6.3.3

A fatal error has been detected by the Java Runtime Environment:
SIGBUS (0x7) at pc=0x00007fe91ef6cebc, pid=30321, tid=0x00007fe930c67700

Cloudera services for HDFS and Impala cannot start after applying the patch.

Commonly used Linux distributions are shown in the table below. However, the issue affects any CDH release that runs on RHEL, CentOS, Oracle Linux, SUSE Linux, or Ubuntu and that has had the Linux kernel security patch for CVE-2017-1000364 applied.

If you have already applied the patch for your OS according to the advisories for CVE-2017-1000364, apply the kernel update that contains the fix for your operating system (some of which are listed in the table). If you cannot apply the kernel update, you can workaround the issue by increasing the Java thread stack size as detailed in the steps below.

Impact: After a leap-second event, Java applications (including CDH services) using older Java and Linux kernel versions, may consume almost 100% CPU. See https://access.redhat.com/articles/15145.

Leap-second events are tied to the time synchronization methods of the Linux kernel, the Linux distribution and version, and the Java version used by applications running on affected kernels.

Although Java is increasingly agnostic to system clock progression (and less susceptible to a kernel's mishandling of a leap-second event), using JDK 7 or 8 should prevent issues at the CDH level (for CDH components that use the Java Virtual Machine).

Immediate action required:

(1) Ensure that the kernel is up to date.

• RHEL6/7, CentOS 6/7 - 2.6.32-298 or higher

• Oracle Enterprise Linux (OEL) - Kernels built in 2013 or later

• SLES12 - No action required.

(3) Ensure that your systems use either NTP or PTP synchronization.

For systems not using time synchronization, update both the OS tzdata and Java tzdata packages to the tzdata-2016g version, at a minimum. For OS tzdata package updates, contact OS support or check updated OS repositories. For Java tzdata package updates, see Oracle's Timezone Updater Tool.

Cloudera Issue: CDH-44788, TSB-189

If an encrypted file channel is set to use fast replay, the replay will fail and the channel will fail to start.

Workaround: Disable fast replay for the encrypted channel by setting use-fast-replay to false.

Apache Issue: FLUME-1885

Several Hadoop and HDFS properties have been deprecated as of Hadoop 3.0 and later. For details, see Deprecated Properties.

Navigator Audit logging for HBase access can have a big impact on HBase performance most noticeable during data ingestion.

Component affected: HBase

Products affected: CDH

Releases affected: CDH 6.x

Impact: 4x performance increase was observed in batchMutate calls after disabling Navigator Audit.

Severity: High

Upgrade: Upgrade to CDP where Navigator is no longer used.

HBASE-25206 can cause data loss either through corrupting an existing hbase snapshot or destroying data that backs a clone of a previous snapshot.

Component affected: HBase

Users affected: Users of the affected releases.

Impact: Potential risk of Data Loss.

Severity: High

Apache issue: HBASE-25206

KB article: For the latest update on this issue see the corresponding Knowledge article: TSB 2021-453: HBASE-25206 "snapshot and cloned table corruption when original table is deleted"

The HDFS short-circuit setting dfs.client.read.shortcircuit is overwritten to disabled by hbase-default.xml. HDFS short-circuit reads bypass access to data in HDFS by using a domain socket (file) instead of a network socket. This alleviates the overhead of TCP to read data from HDFS which can have a meaningful improvement on HBase performance (as high as 30-40%).

Users can restore short-circuit reads by explicitly setting dfs.client.read.shortcircuit in HBase configuration via the configuration management tool for their product (e.g. Cloudera Manager or Ambari).

Impact: HBase reads with high data-locality will not execute as fast as previously. HBase random read performance is heavily affected as random reads are expected to have low latency (e.g. Get, Multi-Get). Scan workloads would also be affected, but may be less impacted as latency of scans is greater.

Severity: High

After making these configuration changes, restart the HBase service.

Cloudera will continue to pursue product changes which may alleviate the need to make these configuration changes.

For CDP 7.1.1.0 and newer, the metric shortCircuitBytesRead can be viewed for each RegionServer under the RegionServer/Server JMX metrics endpoint. When short circuit reads are not enabled, this metric will be zero. When short circuit reads are enabled and the data locality for this RegionServer is greater than zero, the metric should be greater than zero.

Knowledge article: For the latest update on this issue see the corresponding Knowledge article: TSB 2021-463: HBase Performance Issue

HDP and CDH releases suffer from low compaction throughput limits, which cause storefiles to back up faster than compactions can re-write them. This was originally identified upstream in HBASE-21000.

Users affected: Users of above mentioned HDP and CDH versions.

Severity: Medium

Impact: For non-read-only workloads, this will eventually cause back-pressure onto new writes when the blocking store files limit is reached.

Apache issue: HBASE-21000

Knowledge article: For the latest update on this issue see the corresponding Knowledge article: Cloudera Customer Advisory: Default limits for PressureAwareCompactionThroughputController are too low

Cloudera Manager does not allow to deploy multiple HBase services on the same host of an HDFS cluster as by design a DataNode can only have a single HBase service per host. It is possible to have two HBase services on the same HDFS cluster but they have to be on different DataNodes, meaning that there will be one RegionServer per DataNode per HBase cluster. However, that requires additional configuration, for example you have to pin /hbase_enc and /hbase to avoid the HDFS balancer to cluster. However, that requires additional configuration, for example you have to pin /hbase_enc and /hbase to avoid the HDFS balancer to cause issues with data locality.

If Cloudera Manager is not used, you can manage multiple configurations per host for different RegionServers that are part of different HBase clusters but that can lead to multiple issues and difficult troubleshooting procedures. Thus, Cloudera does not support managing multiple HBase services on the same CDH cluster.

CDH 5.12.0 includes the fix HBASE-16604, where the internal scanner that retries in case of IOException from timeouts could potentially miss data. Java clients were properly updated to account for the new behavior, but thrift clients will now see exceptions where the previous missing data would be.

Workaround: Create a new scanner and retry the operation when encountering this issue.

During IntegrationTestReplication, if the verify phase starts before the replication phase finishes, the test will fail because the target cluster does not contain all of the data. If the HBase services in the target cluster does not have enough memory, long garbage-collection pauses might occur.

Workaround: Use the -t flag to set the timeout value before starting verification.

Cloudera Issue: None.

Cloudera has tested the performance impact of using HDFS encryption with HBase. The overall overhead of HDFS encryption on HBase performance is in the range of 3 to 4% for both read and update workloads. Scan performance has not been thoroughly tested.

ExportSnapshot or DistCP operations may fail on AWS when using certain JDK 8 versions, due to an incompatibility between the AWS Java SDK 1.9.x and the joda-time date-parsing module.

Workaround: Use joda-time 2.8.1 or higher, which is included in AWS Java SDK 1.10.1 or higher.

Cloudera Issue: None.

There are two distinct causes for the regressions, depending on the workload:

• For a cached workload, the regression may be up to 11%, as compared to RHEL6. The cause relates to differences in the CPU's C-state (power saving state) behavior. With the same workload, the CPU is around 40% busier in RHEL7, and the CPU spends more time transitioning between C-states in RHEL7. Transitions out of deeper C-states add latency. When CPUs are configured to never enter a C-state lower than 1, RHEL7 is slightly faster than RHEL6 on the cached workload. The root cause is still under investigation and may be hardware-dependent.
• For an IO-bound workload, the regression may be up to 8%, even with common C-state settings. A 6% difference in average disk service time has been observed, which in turn seems to be caused by a 10% higher average read size at the drive on RHEL7. The read sizes issued by HBase are the same in both cases, so the root cause seems to be a change in the EXT4 filesystem or the Linux block IO later. The root cause is still under investigation.

Bug: None

Severity: Medium

Workaround: Avoid using RHEL 7 if you have a latency-critical workload. For a cached workload, consider tuning the C-state (power-saving) behavior of your CPUs.

CDH 5.3 and higher supports Azure Blob Storage for some applications. However, a null pointer exception occurs when you specify a wasb:// or wasbs:// location in the --copy-to option of the ExportSnapshot command or as the output directory (the second positional argument) of the Export command.

Workaround: None.

Apache Issue: HADOOP-12717

When security and Access Control are enabled, the following problems occur:

• If a Delete Table fails for a reason other than missing permissions, the access rights are removed but the table may still exist and may be used again.

• If hbaseAdmin.modifyTable() is used to delete column families, the rights are not removed from the Access Control List (ACL) table. The postOperation is implemented only for postDeleteColumn().

• If Create Table fails, full rights for that table persist for the user who attempted to create it. If another user later succeeds in creating the table, the user who made the failed attempt still has the full rights.

Workaround: None

Apache Issue: HBASE-6992

There are no notable known issues in this release of HCatalog.

Products affected: Hue

User affected: All users of the affected versions

Severity (Low/Medium/High): Medium

Impact:Security Vulnerabilities as mentioned in the CVEs

Issuing DDL statements using Hue’s Impala editor or invoking Hue’s “Refresh Cache” function in the left-side metadata browser results in Hue issuing INVALIDATE METADATA calls to the Impala service. This call is expensive and can result in a significant system impact, up to and including full system outage, when repeated in sufficient volume. This has been corrected in HUE-8882.

Users affected: End-users using Impala editor in Hue.

Severity: High

Impact: Users running DDL statements using the Hue Impala editor or invoking Hue’s Refresh Cache function causes INVALIDATE METADATA commands to be sent to Impala. Impala’s metadata invalidation is an expensive operation and could cause impact on the performance of subsequent queries, hence leading to the potential for significant impact on the entire cluster, including the potential for whole-system outage.

Apache issue: HUE-8882

Knowledge article: For the latest update on this issue see the corresponding Knowledge article: Cloudera Customer Advisory: High DDL usage in Hue Impala Editor may issue flood of INVALIDATE Calls

The Hue Load Balancer merges the double slashes (//) in the S3 URI into a single slash (/) so that the URI prefix "/filebrowser/view=S3A://" is changed to "/filebrowser/view=S3A:/". This results in an error when you try to access the S3 buckets from the Hue File Browser through the port 8889.

The Hue web UI displays the following error: “Unknown error occurred”.

The Hue server logs record the “ValueError: Invalid S3 URI: S3A” error.

Workaround:

You may see an error such as the following while rerunning an an already executed and finished Oozie workflow through the Hue web interface: E0504: App directory [hdfs:/cdh/user/hue/oozie/workspaces/hue-oozie-1571929263.84] does not exist.

Workaround:

When you create a new table from a file by using the Data Import Wizard, the newly created table columns do not display in the Table Browser until you refresh it.

For example:

1. Create a table from the sample file web_logs_2.csv by clicking the plus sign in the left panel, which launches the Data Import Wizard:

   
   

   

   
   

2. After you define the table, click Submit to generate the new table:

   
   

   

   
   

3. After you generate the new table, you can see it listed on the left assist panel, but when you click the table name to display the columns, an error displays:

   
   

   

   
   

Workaround:

1. Click the information icon that is adjacent to the new table:

   
   

   

   
   

2. In the information window that opens, click the refresh icon in the upper right corner to view the table columns:

   
   

   

   
   

   Using the refresh icon in the information window is the least expensive way to refresh the page so performance is not affected.

Affected Version(s): CDH 6.2.0

Cloudera Issue:CDH-77238

Hue does not currently support the Spark application.

These issues can prevent one or more Impala-related daemons from starting properly.

These issues can cause Impala to quit or become unresponsive.

These issues involve the performance of operations such as queries or DDL statements.

These issues relate to security features, such as Kerberos authentication, Sentry authorization, encryption, auditing, and redaction.

These issues involve memory or disk usage, including out-of-memory conditions, the spill-to-disk feature, and resource management features.

These issues can cause incorrect or unexpected results from queries. They typically only arise in very specific circumstances.

These issues affect the ability to interchange data between Impala and other systems. They cover areas such as data types and file formats.

These issues are current limitations of Impala that require evaluation as you plan how to integrate Impala into your data management workflow.

These issues do not fall into one of the above categories or have not been categorized yet.

Topics that are created and deleted via Kafka are secured (for example, auto created topics). However, most topic creation and deletion is done via the kafka-topics tool, which talks directly to ZooKeeper or some other third-party tool that talks directly to ZooKeeper. Because security is the responsibility of ZooKeeper authorization and authentication, Kafka cannot prevent users from making ZooKeeper changes. Anyone with access to ZooKeeper can create and delete topics. They will not be able to describe, read, or write to the topics even if they can create them.

The offsets.topic.replication.factor broker configuration is now enforced upon auto topic creation. Internal auto topic creation will fail with a GROUP_COORDINATOR_NOT_AVAILABLE error until the cluster size meets this replication factor requirement.

The first few produce requests fail when sending to a nonexistent topic with auto.create.topics.enable set to true.

Workaround: Increase the number of retries in the Producer configuration setting retries.

When using ZooKeeper authentication and a custom Kerberos principal, Kerberos-enabled Kafka does not start.

Workaround: None. You must disable ZooKeeper authentication for Kafka or use the default Kerberos principals for ZooKeeper and Kafka.

Significant performance degradation can occur when SSL is enabled. The impact varies depending on your CPU, JVM version, and message size. Consumers are typically more affected than producers.

Workaround: Configure brokers and clients with ssl.secure.random.implementation = SHA1PRNG. It often reduces this degradation drastically, but its effect is CPU and JVM dependent.

Affected Versions: CDK 2.x and later

Fixed Versions: None

Apache Issue: KAFKA-2561

Cloudera Issue: None

When MirrorMaker is used in conjunction with Sentry, MirrorMaker reports an authorization issue and does not start. This is due to Sentry being unable to authorize the kafka_mirror_maker principal which is automatically created.

Affected Versions: CDH 6.0.0 and later

Fixed Versions: N/A

Apache Issue: N/A

Cloudera Issue: CDH-53706

DDL and DML operations can accumulate in the Kudu tablet replica's write ahead log (WAL) during normal operation. Upon the shutdown of a tablet replica (for example, right before removing the replica), information on the accumulated operations (first 50) are printed into the tablet server's INFO log file.

A bug was introduced with the fix for KUDU-2690. The code contains a flipped if-condition that results in de-referencing of an invalid pointer while reporting on a pending ALTER TABLE operation in the tablet replica's WAL. The issue manifests itself in kudu-tserver processes crashing with SIGSEGV (segmentation fault).

The occurrence of the issue is limited to scenarios which result in accumulating at least one pending ALTER TABLE operation in the tablet replica's WAL at the time when the tablet replica is shut down. An example scenario is an ALTER TABLE request (for example, adding a column) immediately followed by a request to drop a tablet (for example, drop a range partition). Another example scenario is shutting down a tablet server while it's still processing an ALTER TABLE request for one of its tablet replicas. A slowness in file system operations increases the chances for the issue to manifest itself.

Apache issue: KUDU-2690

Component affected: Kudu

Products affected: CDH

Users affected: Kudu clusters with the impacted releases.

Impact:In the worst case, multiple kudu-tserver processes can crash in a Kudu cluster, making data unavailable until the affected tablet servers are started back.

Severity: High

Knowledge article: For the latest update on this issue see the corresponding Knowledge article:

TSB 2020-449: Kudu tablet server might crash in certain workflows where a tablet is dropped right after ALTER TABLE statement

If you are using fine grained authorization for Kudu, and you are also using Kudu-HMS integration with HDFS-Sentry sync, then you may get the "The user 'kudu' is not part of group 'hive' on the following hosts: " warning while upgrading to CDH 6.3.

usermod -aG hive kudu

Affected Versions: CDH 6.3 / Kudu 1.10

If a tablet server has a very large number of tablets, it may take several minutes to start up. It is recommended to limit the number of tablets per server to 1000 or fewer. The maximum allowed number of tablets is 2000 per server. Consider this limitation when pre-splitting your tables. If you notice slow start-up times, you can monitor the number of tablets per server in the web UI.

Affected Versions: All CDH 6 versions

Use the descriptions in the following table to better understand the TLS/SSL settings in the Cloudera Manager Admin Console.

Affected Versions: All CDH 6 versions

If the JobHistory server is down on a YARN (MRv2) cluster, Oozie attempts to submit a job, by default, three times. If the job fails, Oozie automatically puts the workflow in a SUSPEND state.

Workaround: When the JobHistory server is running again, use the resume command to tell Oozie to continue the workflow from the point at which it left off.

Affected Versions: CDH 5 and higher

Cloudera Issue: CDH-14623

When performing a shard split on an index that is stored on HDFS, SplitShardCmd still evaluates free disk space on the local file system of the server where Solr is installed. This may cause the command to fail, perceiving that there is no adequate disk space to perform the shard split.

Workaround: None

Affected versions: All

Although it is technically possible to give user-defined Solr core names during core creation, it is to be avoided in te context of Cloudera Search. Cloudera Manager expects core names in the default "collection_shardX_replicaY" format. Altering core names results in Cloudera Manager being unable to fetch Solr metrics for the given core and this, eventually, may corrupt data collection for co-located core, or even shard and server level charts.

The Solr SQL, Graph, and Stream handlers do not support document-level security, and are disabled if document-level security is enabled on the collection. If necessary, these handlers can be re-enabled by setting the following Java system properties, but document-level security is not enforced for these handlers:

• SQL: solr.sentry.enableSqlQuery=true
• Graph: solr.sentry.enableGraphQuery=true
• Stream: solr.sentry.enableStreams=true

Workaround: None

Affected Versions: All CDH 6 releases

Cloudera Issue: CDH-66345

Before CDH 5.5.0, a collection could be created without specifying a configuration. If no -c value was specified, then:

• If there was only one configuration, that configuration was chosen.
• If the collection name matched a configuration name, that configuration was chosen.

Search for CDH 5.5.0 includes multiple built-in configurations. As a result, there is no longer a case in which only one configuration can be chosen by default.

Workaround: Explicitly specify the collection configuration to use by passing -c <configName> to solrctl collection --create.

Affected Versions: CDH 5.5.0 and higher

Cloudera Issue: CDH-34050

If the --input-file-format option is specified with CrunchIndexerTool, then its argument must be text, avro, or avroParquet, rather than a fully qualified class name.

Workaround: None

Affected Versions: All

Cloudera Issue: CDH-22190

The quickstart.sh file uses the timeout function to determine if ZooKeeper and the NameNode are available. To ensure this check can be complete as intended, the quickstart.sh determines if the operating system on which the script is running supports timeout. If the script detects that the operating system does not support timeout, the script continues without checking if the NameNode and ZooKeeper are available. If your environment is configured properly or you are using an operating system that supports timeout, this issue does not apply.

Workaround: This issue only occurs in some operating systems. If timeout is not available, the quickstart continues and final validation is always done by the MapReduce jobs and Solr commands that are run by the quickstart.

Affected Versions: All

Cloudera Issue: CDH-19923

The MapReduceIndexerTool and the HBaseMapReduceIndexerTool can be used with a Managed Schema created via NRT indexing of documents or via the Solr Schema API. However, neither tool supports adding fields automatically to the schema during ingest.

Workaround: Define the schema before running the MapReduceIndexerTool or HBaseMapReduceIndexerTool. In non-schemaless mode, define in the schema using the schema.xml file. In schemaless mode, either define the schema using the Solr Schema API or index sample documents using NRT indexing before invoking the tools. In either case, Cloudera recommends that you verify that the schema is what you expect using the List Fields API command.

Affected Versions: All

Cloudera Issue: CDH-26856

The Browse and Spell Request Handlers require certain fields be present in the schema. Since those fields cannot be guaranteed to exist in a Schemaless setup, the Browse and Spell Request Handlers are not enabled by default.

Workaround: If you require the “Browse” and “Spell” Request Handlers, add them to the solrconfig.xml configuration file. Generate a non-schemaless configuration to see the usual settings and modify the required fields to fit your schema.

Affected Versions: All

Cloudera Issue: CDH-19407

It is possible to create indexes with solr.hdfs.blockcache.write.enabled set to true. Such indexes may appear corrupt to readers, and reading these indexes may irrecoverably corrupt indexes. Blockcache writing is disabled by default.

Workaround: None

Affected Versions: All

Cloudera Issue: CDH-17978

Users who are not authorized to use the Solr Admin UI are not given page explaining that access is denied, and instead receive a web page that never finishes loading.

Workaround: None

Affected Versions: All

Cloudera Issue: CDH-58276

Repeatedly running the MapReduceIndexerTool on the same set of input files can result in duplicate entries in the Solr collection. This occurs because the tool can only insert documents and cannot update or delete existing Solr documents. This issue does not apply to the HBaseMapReduceIndexerTool unless it is run with more than zero reducers.

Workaround: To avoid this issue, use HBaseMapReduceIndexerTool with zero reducers. This must be done without Kerberos.

Affected Versions: All

Cloudera Issue: CDH-15441

It is possible to delete a collection when hosts that host some of the collection are unavailable. After such a deletion, if the previously unavailable hosts are brought back online, the deleted collection may be restored.

Workaround: Ensure all hosts are online before deleting collections.

Affected Versions: All

Cloudera Issue: CDH-58694

Cloudera Search does not support the ability to save search results.

Workaround: None

Affected Versions: All

Cloudera Issue: CDH-21162

Cloudera Search does not support HDFS Federation.

Workaround: None

Affected Versions: All

Cloudera Issue: CDH-11357

Solr contrib modules are not supported (Morphlines, Spark Crunch indexer, MapReduce and Lily HBase indexers are part of the Cloudera Search product itself, therefore they are supported).

Workaround: None

Affected Versions: All

Cloudera Issue: CDH-72658

Using the Sentry Service with Cloudera Search may introduce latency because authorization requests must be sent to the Sentry Service.

Workaround: You can alleviate this latency by enabling caching for the Sentry Service. For instructions, see: Enabling Caching for the Sentry Service.

Affected Versions: All

Cloudera Issue: CDH-73407

If multiple Solr instances are configured to depend on the same Sentry service, it is not possible to create unique Solr Sentry privileges per Solr deployment. Since privileges are enforced in all Solr instances simultaneously, you cannot add distinct privileges that apply to one Solr cluster, but not to another.

Workaround: None

Affected Versions: All

Cloudera Issue: CDH-72676

If you enable TLS/SSL on a Solr instance with existing collections, the collections will break and become unavailable. Collections created after enabling TLS/SSL are not affected by this issue.

Workaround: Recreate the collection after enabling TLS. For more information, see How to update existing collections in Non-SSL to SSL in Solr.

Affected Versions: All

Cloudera Issue: CDPD-4139

A Kafka topic name can have 249 characters, but Sentry only supports topic names up to 64 characters.

Workaround: Keep Kafka topic names to 64 charcters or less.

Affected Versions: All CDH 5.x and 6.x versions

Cloudera Issue: CDH-64317

If you run the command SHOW ROLE GRANT GROUP for a group that has never been granted a role, beeline raises an exception. However, if you run the same command for a group that does not have any roles, but has at one time been granted a role, you do not get an exception, but instead get an empty list of roles granted to the group.

Workaround: Adding a role will prevent the exception.

Affected Versions:

Cloudera Issue: CDH-71694

Under a significant workload, Grant/Revoke operations can have issues.

Workaround: If you need to make many privilege changes, plan them at a time when you do not need to do too many at once.

Affected Versions: CDH 5.13.0 and above

Apache Issue: SENTRY-1855

Cloudera Issue: CDH-56553

Using more than a thousand roles/permissions might cause significant performance problems.

Workaround: Plan your roles so that groups have as few roles as possible and roles have as few permissions as possible.

Affected Versions: CDH 5.13.0 and above

Cloudera Issue: CDH-59010

As a prerequisite of enabling Sentry, Hive impersonation is turned off, which means all YARN jobs are submitted to the Hive job queue, and are run as the hive user. This is an issue because the YARN History Server now has to block users from accessing logs for their own jobs, since their own usernames are not associated with the jobs. As a result, end users cannot access any job logs unless they can get sudo access to the cluster as the hdfs, hive or other admin users.

In CDH 5.8 (and higher), Hive overrides the default configuration, mapred.job.queuename, and places incoming jobs into the connected user's job queue, even though the submitting user remains hive. Hive obtains the relevant queue/username information for each job by using YARN's fair-scheduler.xml file.

Affected Versions: CDH 5.2.0 and above

Cloudera Issue: CDH-22890

GRANT and REVOKE for column level privileges is not supported on Hive Metastore views.

Affected Versions: All CDH versions

Apache Issue: SENTRY-754

Users who have been explicitly granted the SELECT privilege on all columns of a table, will not have the permission to perform table-level operations. For example, operations such as SELECT COUNT (1) or SELECT COUNT (*) will not work even if you have the SELECT privilege on all columns.

There is one exception to this. The SELECT * FROM TABLE command will work even if you do not have explicit table-level access.

Affected Versions: All CDH versions

Apache Issue: SENTRY-838

Users are able to run the EXPLAIN SELECT operation, exposing metadata for all columns, even for tables/columns to which they weren't explicitly granted access.

Affected Versions: All CDH versions

Apache Issue: SENTRY-849

Workaround:Use SHOW GRANT ROLE roleNameto list all privileges granted to the role.

Affected Versions: All CDH versions

Apache Issue: N/A

Cloudera Issue: CDH-19430

Sentry supports only absolute (not relative) URI paths in permission grants. Although some early releases (for example, CDH 5.7.0) might not have raised explicit errors when relative paths were set, upgrading a system that uses relative paths causes the system to lose Sentry permissions.

Resolution: Revoke privileges that have been set using relative paths, and grant permissions using absolute paths before upgrading.

Affected Versions: All versions. Relative paths are not supported in Sentry for permission grants.

The RDD.repartition() transformation, which reshuffles data in the RDD randomly to create either more or fewer partitions and then balances it across the partitions, was using a round-robin method to distribute data that caused incorrect answers to be returned for RDD jobs. This issue has been corrected, but it introduced a behavior change in RDD job failure handling. Now, Spark actively fails a job if there is a fetch failure that was caused by a node failure after repartitioning.

Workaround: Use the RDD.checkpoint() method to save the intermediate RDD data to HDFS. First, call SparkContext.setCheckpointDir(directory: String) to set the checkpoint directory where the intermediate data will be saved. Note that the directory must be an HDFS path. Then mark the RDD for checkpointing by calling RDD.checkpoint() when you use the RDD.repartition() transformation.

Apache Issue: SPARK-23243

Cloudera Issue: CDH-76413

In Spark Structured Streaming, the exactly-once fault tolerance for file sink is valid only for files that are in the manifest. These files are located in the _spark_metadata subdirectory of the file sink output directory. Only process files that have file names starting with digits. Other temporary files can also appear in this directory, but they should not be processed. Typically, these temporary file file names start with a period (".").

You can list the valid manifest files, excluding the temporary files, by using a command like the following, which assumes your output directory is located at /tmp/output. As the appropriate user, run the following command to list the valid manifest files:

hadoop fs -ls /tmp/output/_spark_metadata/[0-9]*
      

Workaround: None

Affected Versions: CDH 6.1.0 and higher

Cloudera Issue: CDH-75191

Spark SQL treats varchar as a string (that is, there no size limit). The observed behavior is that Spark reads and writes these columns as regular strings; if inserted values exceed the size limit, no error will occur. The data will be truncated when read from Hive, but not when read from Spark.

Workaround: None

Affected Versions: CDH 5.5.0 and higher

Apache Issue: SPARK-5918

Cloudera Issue: CDH-33642

Spark allows you to declare DataFrames with any key type. Avro supports only string keys and trying to write any other key type to an Avro table will fail.

Workaround: None

Affected Versions: CDH 5.5.0 and higher

Cloudera Issue: CDH-33648

Workaround: None

Affected Versions: CDH 5.5.0 and higher

Cloudera Issue: CDH-33649

If you are using Spark Streaming, Cloudera recommends that you disable dynamic allocation by setting spark.dynamicAllocation.enabled to false when running streaming applications.

When SparkSQL accesses an HBase table through the HiveContext, region pruning is not performed. This limitation can result in slower performance for some SparkSQL queries against tables that use the HBase SerDes than when the same table is accessed through Impala or Hive.

Workaround: None

Affected Versions: All

Cloudera Issue: CDH-56330

The spark-submit script's --principal and --keytab arguments do not work with Spark-on-YARN's client mode.

Workaround: Use cluster mode instead.

Affected Versions: All

Using the --proxy-user argument in client mode does not work and is not supported.

Workaround: Use cluster mode instead.

Affected Versions: All

When a Spark application is killed, the history link in the ResourceManager web UI does not work.

Workaround: To view the history for a killed Spark application, see the Spark HistoryServer web UI instead.

Affected Versions: All CDH versions

Apache Issue: None

Cloudera Issue: CDH-49165

Currently, Sqoop is using an Avro schema when writing data as a parquet file. The Avro schema requires that column names do not start with numbers, therefore Sqoop is renaming the columns in this case, prepending them with an underscore character. This can lead to issues when one wants to reuse the data in other tools, such as Impala.

Workaround: Rename the columns to comply with Avro limitations (start with letters or underscore, as specified in the Avro documentation).

Cloudera Issue: None

CentOS 6 systems currently ship with version 5.1.17 of the MySQL JDBC driver. This version does not work correctly with Sqoop.

Workaround: Install version 5.1.31 of the JDBC driver as detailed in Installing the JDBC Drivers for Sqoop 1.

Affected Versions: MySQL JDBC 5.1.17, 5.1.4, 5.3.0

Cloudera Issue: CDH-23180

The integratedSecurity option is not available in the Sqoop CLI.

Workaround: None

Cloudera Issue: None

sqoop import --connect jdbc:db2://djaxludb1001:61035/DDBAT003 --username=dh810202 --P --target-dir /data/hive_scratch/ASDISBURSEMENT --delete-target-dir -m1 --query "select disbursementnumber,disbursementdate,xmldata FROM DB2dba.ASDISBURSEMENT where DISBURSEMENTNUMBER = 2011113210000115311 AND \$CONDITIONS" -hive-import --hive-database adminserver -hive-table asdisbursement_dave --map-column-java XMLDATA=String --as-parquetfile

16/12/05 12:23:46 INFO mapreduce.Job: map 100% reduce 0%
16/12/05 12:23:46 INFO mapreduce.Job: Job job_1480530522947_0096 failed with state FAILED due to: Job commit failed: org.kitesdk.data.DatasetIOException: Could not move contents of hdfs://AJAX01-ns/tmp/adminserver/.temp/job_1480530522947_0096/mr/job_1480530522947_0096 to hdfs://AJAX01-ns/data/RetiredApps/INS/AdminServer/asdisbursement_dave
<SNIP>
Caused by: org.apache.hadoop.ipc.RemoteException(java.io.IOException): /tmp/adminserver/.temp/job_1480530522947_0096/mr/job_1480530522947_0096/5ddcac42-5d69-4e46-88c2-17bbedac4858.parquet can't be moved into an encryption zone.

Workaround: If you use the Parquet Hadoop API based implementation for importing into Parquet, specify a --target-dir which is the same encryption zone as the Hive warehouse directory.

If you use the Kite Dataset API based implementation, use an alternate data file type, for example text or avro.

Apache Issue: SQOOP-2943

Cloudera Issue: CDH-40826

Workaround: None, other than restructuring the data.

Apache Issue: PARQUET-99