What's New In CDH 5.7.x

• Operating Systems - Support for:
  • RHEL/CentOS 6.6, 6.7, 7.1, and 7.2
  • Oracle Enterprise Linux (OEL) 7.1 and 7.2
  • SUSE Linux Enterprise Server (SLES) 11 with Service Packs 2, 3, 4
  • Debian: Wheezy 7.0, 7.1, and 7.8

• Improve support for heterogeneous storage. Heterogeneous Storage Management (HSM) is now supported natively in Cloudera Manager.
• HADOOP-10651 - Service access can be restricted by IP and hostname. You can now define a whitelist and blacklist per service. The default whitelist is * and the default blacklist is empty.
• HADOOP-12764 - You can configure KMS maxHttpHeaderSize. The default value of KMS maxHttpHeaderSize increased from 4096 to 65536 and is now configurable in service.xml.
• HDFS-7279 - In CDH 5.5.0 and higher, DataNode WebHDFS implementation uses Netty as an HTTP server instead of Jetty. With improved buffer and connection management, Netty lowers the risk for DataNode latency and OutOfMemoryError (OOM).
• HDFS-8873 - You can rate-limit the directoryScanner. A new configuration property, dfs.datanode.directoryscan.throttle.limit.ms.per.sec, allows you to reduce the impact on disk performance of directory scanning. The default value is 1000.
• HDFS-9260 - Garbage collection of full block reports is improved. Data structures for BlockInfo and replicas were changed to keep them sorted. This allows for faster and easier garbage collection of full block reports.

See also Apache HBase Known Issues.

• CDH 5.7.0 adds support for a snapshot owner for a table. To configure a table snapshot owner, set the OWNER attribute on the table to a valid HBase user. By default, the table owner is the user who created the table. The table owner and the table creator can restore the table from a snapshot.
• You can set an HDFS storage policy to store write-ahead logs (WALs) on solid-state drives (SSDs) or a mix of SSDs and spinning disks. See Configuring the Storage Policy for the Write-Ahead Log (WAL).
• Configuration for snapshot timeouts has been simplified. A single configuration option, hbase.snapshot.master.timeout.millis, controls how long the master waits for a response from the RegionServer before timing out. The default timeout value has been increased from 60,000 milliseconds to 300,000 to accommodate larger tables.
• You can optionally balance a table's regions by size by setting the hbase.normalizer.enabled property to true. The default value is false. To configure the time interval for the HMaster to check for region balance, set the hbase.normalizer.period property to the interval, in milliseconds. The default value is 1800000, or 30 minutes.
• You can configure parallel request cancellation for multi-get operations using the hbase.client.replica.interrupt.multiget configuration property, using an advanced configuration snippet in Cloudera Manager, or in hbase-site.xml if you do not use Cloudera Manager.
• The hbase-spark module has been added, which provides support for using HBase data in Spark jobs using HBaseContext and JavaHBaseContext contexts. See the HBase and Spark chapter of the Apache HBase Reference Guide for details about building a Spark application with HBase support.
• The REST API now supports creating, reading, updating, and deleting namespaces.
• If you use the G1 garbage collector, you can disable the BoundedByteBufferPool. See Disabling the BoundedByteBufferPool.
• The HBase web UI includes graphical tools for managing MemStore and StoreFile details for a region.
• The HBase web UI displays the number of regions a table has.
• Two new methods of the Scan API, setColumnFamilyTimeRange and getColumnFamilyTimeRange, allow you to limit Scan results to versions of columns within a specified timestamp range.
• A new API, MultiRowRangeFilter, allows you to scan multiple row ranges in a single scan. See MultiRowRangeFilter.
• A new client API, SecurityCapability, enables you to check whether the HBase server supports cell-level security.
• When using the scan command in HBase Shell, you can use the ROWPREFIXFILTER option to include only rows matching a given prefix, in addition to any other filters you specify. For example:

  hbase> scan 't1', {ROWPREFIXFILTER => 'row2',
                           FILTER => (QualifierFilter (>=, 'binary:xyz')) AND (TimestampsFilter ( 123, 456))}

• The new get_splits HBase Shell command returns the split points for a table. For example:

  hbase> get_splits 't2'
  Total number of splits = 5
  
  => ["", "10", "20", "30", "40"]

• Three new commands relating to the region nomalizer have been added to the HBase Shell:
  • normalizer_enabled checks whether the region normalizer is enabled.
  • normalizer_switch toggles the region normalizer on or off.
  • normalize runs the region normalizer if it is enabled.
• When configuring a ZooKeeper quorum for HBase, you can now specify the port for each ZooKeeper host separately, instead of using the same port for each ZooKeeper host. The configuration property hbase.zookeeper.clientPort is no longer required. For example:

  <property>
    <name>hbase.zookeeper.quorum</name>
    <value>zk1.example.com:2181,zk2.example.com:20000,zk3.example.com:31111</value>
  </property>

• A new configuration option, hbase.regionserver.hostname, was added, but Cloudera recommends against its use. See Compatibility Notes for CDH 5.7.
• Two new configuration options allow you to disable loading of all coprocessors or only table-level coprocessors: hbase.coprocessor.enabled (which defaults to true) and hbase.coprocessor.user.enabled (which also defaults to true). Cloudera does not recommend disabling HBase-wide coprocessors, because security functionality is implemented using coprocessors. However, disabling table-level coprocessors may be appropriate in some cases. See Disabling Loading of Coprocessors.
• A new configuration option, hbase.loadincremental.validate.hfile, allows you to skip validation of HFiles during a bulk load operation when set to false. The default setting is true.
• The default PermSize for HBase processes is now set to 128 MB. This setting is effective only for JDK 7, because JDK 8 ignores the PermSize setting. To change this setting, edit the HBase Client Environment Advanced Configuration Snippet (Safety Valve) for hbase-env.sh if you use Cloudera Manager or conf/hbase-env.sh otherwise.
• In CDH 5.6 and lower, the HBaseFsck#checkRegionConsistency() method would throw an IOException if a region repair operation timed out after hbase.hbck.assign.timeout (which defaults to 120 seconds). This exception would the entire hbck operation to fail. In CDH 5.7.0, if the region being repaired is not hbase:meta or another system table, the region is skipped, an error is logged, and the hbck operation continues. This new behavior is disabled by default; to enable it, set the hbase.hbck.skipped.regions.limit option to an integer greater than 0. If more than this number of regions is skipped, the hbck operation fails.
• Two new options, -exclusive and -disableBalancer, have been added to the hbck utility. The hbck utility now runs without locks unless in fixer mode, and the balancer is only disabled in fixer mode, by default. You can disable these options to retain the old behavior, but Cloudera recommends using the new default behavior.
• Two new MapReduce jobs, SyncTable and HashTable, allow you to synchronize two different HBase tables that are each receiving live writes. To print usage instructions, run the job with no arguments. These examples show how to run these jobs:

  $ bin/hbase org.apache.hadoop.hbase.mapreduce.SyncTable \
                                              --dryrun=true \
                                              --sourcezkcluster=zk1.example.com,zk2.example.com,zk3.example.com:2181:/hbase \
                                              hdfs://nn:9000/hashes/tableA \
                                              tableA tableA

  $ bin/hbase org.apache.hadoop.hbase.mapreduce.HashTable \
                                              --batchsize=32000 \
                                              --numhashfiles=50 \
                                              --starttime=1265875194289 \
                                              --endtime=1265878794289 \
                                              --families=cf2,cf3 \
                                              TestTable /hashes/testTable

• The CopyTable command now allows you to override the org.apache.hadoop.hbase.mapreduce.TableOutputFormat property by prefixing the property keys with the hbase.mapred.output. prefix. For example, hbase.mapred.output.hbase.security.authentication is passed to CopyTable as hbase.security.authentication. This is useful when directing output to a peer cluster with different configuration settings.
• Several improvements have been made to the HBase canary, including:
  • Sniffing of regions and RegionServers is now parallel to improve performance in large clusters with more than 1000 regions and more than 500 RegionServers.
  • The canary sets cachecheck to false when performing Gets and Scans to avoid influencing the BlockCache.
  • FirstKeyOnlyFilter is used during Gets and Scans to improve performance in a flat wide table.
  • A region is selected at random when sniffing a RegionServer.
  • The sink class used by the canary is now configurable.
  • A new flag, -allRegions, sniffs all regions on a RegionServer if running in RegionServer mode.
• Distributed log replay has been disabled in CDH 5.7 and higher, due to reliability issues. Distributed log replay is unsupported.
• A new configuration option, hbase.hfile.drop.behind.compaction, causes the OS-level filesystem cache to be dropped behind compactions. This provides significant performance improvements on large compactions. To disable this behavior, set the option to false. It defaults to true.
• Two new configuration options, hbase.hstore.compaction.max.size and hbase.hstore.compaction.max.size.offpeak, allow you to specify a maximum compaction size during normal hours and during off-peak hours, if the off-peak feature is used. If unspecified, hbase.hstore.compaction.max.size defaults to 9,223,372,036,854,775,807 (Long.MAX_VALUE), which is essentially unbounded.
• The HDFS replication factor can now be specified per column family by setting the DFS_REPLICATION attribute of the column family to the desired number of replicas, either at table creation or by altering an existing table schema. If set to 0, the default replication factor is used. If fewer than the desired number of replicas exist, the HDFS FSCK utility reports it.
• If you use the ChaosMonkey tool, you can low-load a custom ChaosMonkey implementation by passing the class to the -m or --monkey option of the ChaosMonkey tool, in the same way that you would normally pass SLOW or CALM.
• A new configuration option, hbase.use.dynamic.jars, allows you to disable the dynamic classloader if set to false.

• HIVE-9298 - Supports reading alternate timestamp formats. The SerDe property, timestamp.formats, is added to allow you to pass in a comma-delimited list of alternate timestamp formats. For example, the following ALTER TABLE statement (in this case, with Joda date-time parsing) adds: yyyy-MM-dd'T'HH:mm:ss and milliseconds since Unix epoch, represented by the special case pattern millis.

  ALTER TABLE timestamp_formats SET SERDEPROPERTIES ("timestamp.formats"="yyyy-MM-dd'T'HH:mm:ss,millis");

• HIVE-7292 - Hive on Spark (HoS) General Availability. This release supports running Hive jobs using Spark as an additional execution backend. This improves performance for Hive queries, especially those involving multiple reduce stages. See Running Apache Hive on Spark in CDH.
• HIVE-12338 - Add Web UI to HiveServer2. Exposes a Web UI on the HiveServer2 service to surface process metrics (JVM stats, open session, and similar) and per-query runtime information (query plan, performance logs). The Web UI is available on port 10002 by default. See Using HiveServer2 Web UI in CDH.
• HIVE-10115 - Handle LDAP and Kerberos authentication on the same HiveServer2 interface. In a Kerberized cluster when alternate authentication is enabled on HiveServer2, it accepts Kerberos authentication. Before this enhancement, when LDAP authentication to HiveServer2 was enabled, the service blocked acceptance of other authentication mechanisms.
• Hive Metastore / HiveServer2 metadata scalability improvements. Fixes to improve scalability and performance include support for DirectSQL by default for metastore operations and incremental memory usage improvements.
• HIVE-12271 - Add additional catalog and query execution metrics for Hive Metastore / HiveServer2 and integrate with Cloudera Manager. Additional Hive metrics include catalog information (number of databases, tables, and partitions in Hive Metastore) and query execution stats (number of worker threads used, job submission times, and statistics for planning and compilation times). All metrics are available in Cloudera Manager.

• Hive metastore service improvements make browsing Hive data faster and easier.
• SQL editor improvements:
  • SQL assist scales to thousands of tables and databases.
  • Improved row and column headers.
  • Button to delete the query history.
  • Button to format SQL.
• Security improvements:
  • Timeout property for idle sessions. See Idle Session Timeout on the Hue Security page.
  • Customizable splash screen on the login page.
  • Support for password-protected SAML certificates.
  • Custom xmlsec1 binary for multiple SAML protocols.
  • Ability to synchronize groups on login for any authentication backend.
• Oozie application improvements:
  • Display graph of external workflow.
  • Option to dry run jobs on submission.
  • User timezone recognition.
  • Automatic email on failure.
  • Ability to execute individual actions.

See New Features in Impala 2.5.x / CDH 5.7.x.

• The mapred job -history command provides an efficient way to fetch history for MapReduce jobs.

• OOZIE-2411 - The Oozie email action now supports BCC:, as well as To: and CC:.

• Improvements to loadSolr. For additional information about these improvements, see the Morphlines Reference Guide. Improvements include:
  • loadSolr retries SolrJ requests if exceptions, such as connection resets, occur. This is useful in cases such as temporary Solr server overloads or transient connectivity failures. This behavior is enabled by default, but can be disabled by setting the Java system property org.kitesdk.morphline.solr.LoadSolrBuilder.disableRetryPolicyByDefault=true.
  • loadSolr includes an optional rate-limiting parameter that is used to set the maximum number of morphline records to ingest per second. The default value is no limit.

• Spark is rebased on Apache Spark 1.6.0.
• SPARK-10000 - Spark 1.6.0 includes a new unified memory manager. The new memory manager is turned off by default (unlike Apache Spark 1.6.0), to make it easier for users to migrate existing workloads, but it is supported.
• SPARK-9999 - Spark 1.6.0 introduces a new Dataset API. However this API is experimental, likely to undergo some changes, and unsupported.
• SPARK-6028 and SPARK-6230 - Encryption support for Spark RPC
• SPARK-2750 - HTTPS support for History Server and web UI
• Added support for Spark SQL (DataFrames) in PySpark.
• Added support for the following MLlib features:
  • spark.ml
  • ML pipeline APIs
• The hbase-spark module has been added, which provides support for using HBase data in Spark jobs using HBaseContext and JavaHBaseContext contexts. See the HBase and Spark chapter of the Apache HBase Reference Guide for details about building a Spark application with HBase support.
• The DATE datatype is supported for Parquet and Avro tables.

SENTRY-510 - Added support for collection of metrics for the Sentry/HDFS plugin