Introduction 

Phase 1 of replication is to support a hot-cold setup where CDAP data is replicated from one cluster to another using existing tools for replicating underlying infrastructure.

Goals

Allow manual failover from a hot cluster to a cold cluster.

User Stories 

• As a cluster administrator, I want to be able to configure CDAP so that all HBase tables created by CDAP are set up to replicate data to another cluster
• As a cluster administrator, I want to be able to manually stop CDAP in one cluster and start it in another cluster with the exact same state
• As a cluster administrator, I want to be able to have a way to know when it is safe to start the cold cluster after the hot one has been shut down

Design

CDAP stores state in several systems:

HDFS

• Transaction snapshots
• Artifacts (jars)
• Streams
• FileSet based datasets
• Program logs

HBase

• CDAP entity metadata (program specifications, schedules, run history, metrics, etc.)
• Table based datasets
• Kafka offsets for metrics and logs
• Flow queues
• Messaging system data

Kafka

• unprocessed metrics
• unsaved log messages

Hive

• Explorable CDAP datasets and their partitions

For phase 1, much of the responsiblity for data replication falls to the cluster administrator. It is assumed that replication of HDFS, Hive, and Kafka will be handled by the cluster administrator. HDFS is usually done through regularly scheduled distcp jobs, or by using some distro specific tools, such as Cloudera's Backup and Data Recovery (http://www.cloudera.com/documentation/enterprise/latest/topics/cm_bdr_about.html). Kafka can be done using MirrorMaker. Hive can be done by replicating the data (HDFS and/or HBase), and by replication the metastore through whatever replication mechanisms are available to the relational DB backing the metastore. All of this can be setup outside of CDAP.

HBase DDL

HBase, however, will require some hooks in CDAP, because replication must be setup for every table when it is created, and before any data is written to it. CDAP will define an interface to create, modify, and delete HBase tables. Instead of just creating a table in the local HBase instance, we need to create a table in both the master and slave instances and set up replication from the master to the slave. We can do this by introducing an SPI for HBase DDL operations, where the default implementation is the current single cluster implementation, and users can plug in their own implementation that creates tables and sets up replication as needed.

Java SPI

/**
 * Executes HBase DDL operations.
 */
public interface HBaseDDLExecutor {

  /**
   * Create the specified namespace if it does not exist.
   *
   * @param name the namespace to create
   * @throws IOException if a remote or network exception occurs
   */
  void createNamespaceIfNotExists(String name) throws IOException;

  /**
   * Delete the specified namespace if it exists.
   *
   * @param name the namespace to delete
   * @throws IOException if a remote or network exception occurs
   * @throws IllegalStateException if there are tables in the namespace
   */
  void deleteNamespaceIfExists(String name) throws IOException;

  /**
   * Create the specified table if it does not exist.
   *
   * @param descriptor the descriptor for the table to create
   * @param splitKeys
   * @throws IOException if a remote or network exception occurs
   * @throws NotFoundException if the namespace for the specified table does not exist
   */
  void createTableIfNotExists(HTableDescriptor descriptor, byte [][] splitKeys) throws IOException;

  /**
   * Enable the specified table.
   *
   * @param name the table to enable
   * @throws IOException if a remote or network exception occurs
   * @throws NotFoundException if the specified table does not exist
   */
  void enableTable(TableName name) throws IOException;

  /**
   * Disable the specified table.
   *
   * @param name the table to disable
   * @throws IOException if a remote or network exception occurs
   * @throws NotFoundException if the specified table does not exist
   */
  void disableTable(TableName name) throws IOException;

  /**
   * Modify the specified table.
   *
   * @param name the table to modify
   * @param descriptor the descriptor for the table
   * @throws IOException if a remote or network exception occurs
   * @throws NotFoundException if the specified table does not exist
   */
  void modifyTable(TableName name, HTableDescriptor descriptor) throws IOException;
 
  /**
   * Truncate the specified table.
   *
   * @param name the table to truncate
   * @throws IOException if a remote or network exception occurs
   * @throws NotFoundException if the specified table does not exist
   */
  void truncateTable(TableName name) throws IOException;

  /**
   * Delete the table if it exists.
   *
   * @param name the table to delete
   * @throws IOException if a remote or network exception occurs
   * @throws NotFoundException if the namespace for the specified table does not exist
   */
  void deleteTableIfExists(TableName name) throws IOException;
}

The default implementation will simply use the existing HBaseTableUtil. There can be another implementation that makes REST calls for each method, leaving actual HBase operations and auth up to an external service. For example, an analagous RESTful API could be:

where the user is passed as request headers. Each endpoint must be idempotent, as there could be a failure in one or more HBase instances, but a success in another instance. In such cases, the client will retry the request, so the endpoint must be idempotent. A 200 should only be returned if the operation succeeded in all HBase instances.

Coprocessors

One difficulty will be in handling the coprocessor jar. Today, when a Table is being created, its coprocessor jar is also built and placed on HDFS.

One way to handle this is to send the jar contents as part of the table creation request (Base64 encoded for example). However, this would be an issue if the master and slave clusters are running different versions of HBase, which require different coprocessors. In order for it to work, we would have to somehow consolidate all coprocessors into a single one that works for all supported HBase versions. This type of thing may be required to support resiliency against rolling HBase upgrades anyway.

If that is not possible, an alternative is for each CDAP instance to pre-build these coprocessor jars and place them on HDFS in pre-determined locations. Instead of building the jar on demand, we simply use the correct location. For example, there may be a tool that users have to run that will build coprocessors compatible with the cluster's hbase version and put them on hdfs.One thing CDAP needs to ensure is that there are no cluster specific values in any of the metadata. For example, the namenode should not be in any of the system metadata, otherwise things will fail when the data is replicated over to the slave and the slave is started.

HBase DDL Design

HBase DDL will require some hooks in CDAP, because replication must be setup for every table when it is created, and before any data is written to it. Design details are at HBase DDL SPI.

Replication Status

Cluster administrators will require a way to tell when it is safe for a cold cluster to be started up. In other words, they need to be able to tell when all necessary data has been replicated. HBase shell already includes a command that helps:

hbase(main):030:0> status 'replication', 'source'
version 1.1.2.2.3.4.7-4
1 live servers
    [hostname]:
       SOURCE: PeerID=1, AgeOfLastShippedOp=29312, SizeOfLogQueue=0, TimeStampsOfLastShippedOp=Thu Nov 10 22:51:55 UTC 2016, Replication Lag=29312

HBase also includes a mapreduce job that can be used to verify replicated data (https://hbase.apache.org/book.html#_verifying_replicated_data).  It must be run on the master cluster.

$ HADOOP_CLASSPATH=`hbase classpath` hadoop jar /usr/hdp/current/hbase-master/lib/hbase-server-1.1.2.2.3.4.7-4.jar verifyrep <peer id> <table>
...
	Map-Reduce Framework
		Map input records=1
		Map output records=0
		Input split bytes=103
		Spilled Records=0
		Failed Shuffles=0
		Merged Map outputs=0
		GC time elapsed (ms)=64
		CPU time spent (ms)=1810
		Physical memory (bytes) snapshot=255139840
		Virtual memory (bytes) snapshot=916021248
		Total committed heap usage (bytes)=287309824
	org.apache.hadoop.hbase.mapreduce.replication.VerifyReplication$Verifier$Counters
		BADROWS=1
		CONTENT_DIFFERENT_ROWS=1

Under the HBase counters, you only want to see the GOODROWS counter, and not BADROWS or CONTENT_DIFFERENT_ROWS.

Design details are at Replication Status Tool. 

Kafka offset mismatches

MirrorMaker is not much more than a Kafka client that consumes from source topics and writes the same messages to some destination. As such, partitions and offsets are not guaranteed to be the sameMirrorMaker does not offer any guarantees about a message from the source being written to the same partition and offset in the destination. The log saver , and metrics processor store Kafka offsets per topic partition in HBase, and their corresponding fetch endpoints will need to be able modified to handle the fact that Kafka offsets can be different in the hot and cold clusters.

Approach

Approach #1

Approach #2

a cluster failover. Design details are at Resolving Kafka Offset Mismatches.

API changes

New Programmatic APIs

New Java APIs introduced (both user facing and internal)See HBase DDL SPI

Deprecated Programmatic APIs

No programmatic APIs will be deprecated

New REST APIs

No new REST APIs will be added to the platform. There may be new REST APIs used by an external service used to handle table DDL and replication.

Deprecated REST API

No REST APIs will be deprecated

CLI Impact or Changes

• No Changes

UI Impact or Changes

• No Changes

Security Impact 

What's the impact on Authorization and how does the design take care of this aspectCluster administrators are responsible for setting up replication, but we should understand what is required from a security perspective to replicate hdfs and hbase data.

Impact on Infrastructure Outages 

With replication, there is now another cluster that is required. If the cold cluster suffers an outage, replication will eventually catch up once service is restored. This assumes the outage lasts is for a shorter duration of time than how long HBase keeps events in its WAL.

Test Scenarios

Releases

Release 4.1.0 

Phase 1 work is scheduled for release 4.1.0.

Related Work

• Work #1
• Work #2
• Work #3

Future work