Apache Spark is a software framework that is used to process data in memory in a distributed manner, and is replacing MapReduce in many use cases.
Spark itself is out of scope of this document, please refer to the Spark site for more information on the Spark project and subprojects. This document will focus on 4 main interaction points between Spark and HBase. Those interaction points are:
- Basic Spark
-
The ability to have a HBase Connection at any point in your Spark DAG.
- Spark Streaming
-
The ability to have a HBase Connection at any point in your Spark Streaming application.
- Spark Bulk Load
-
The ability to write directly to HBase HFiles for bulk insertion into HBase
- SparkSQL/DataFrames
-
The ability to write SparkSQL that draws on tables that are represented in HBase.
The following sections will walk through examples of all these interaction points.
1. Basic Spark
This section discusses Spark HBase integration at the lowest and simplest levels. All the other interaction points are built upon the concepts that will be described here.
At the root of all Spark and HBase integration is the HBaseContext. The HBaseContext takes in HBase configurations and pushes them to the Spark executors. This allows us to have an HBase Connection per Spark Executor in a static location.
For reference, Spark Executors can be on the same nodes as the Region Servers or on different nodes there is no dependence of co-location. Think of every Spark Executor as a multi-threaded client application. This allows any Spark Tasks running on the executors to access the shared Connection object.
This example shows how HBaseContext can be used to do a foreachPartition
on a RDD
in Scala:
val sc = new SparkContext("local", "test")
val config = new HBaseConfiguration()
...
val hbaseContext = new HBaseContext(sc, config)
rdd.hbaseForeachPartition(hbaseContext, (it, conn) => {
val bufferedMutator = conn.getBufferedMutator(TableName.valueOf("t1"))
it.foreach((putRecord) => {
. val put = new Put(putRecord._1)
. putRecord._2.foreach((putValue) => put.addColumn(putValue._1, putValue._2, putValue._3))
. bufferedMutator.mutate(put)
})
bufferedMutator.flush()
bufferedMutator.close()
})
Here is the same example implemented in Java:
JavaSparkContext jsc = new JavaSparkContext(sparkConf);
try {
List<byte[]> list = new ArrayList<>();
list.add(Bytes.toBytes("1"));
...
list.add(Bytes.toBytes("5"));
JavaRDD<byte[]> rdd = jsc.parallelize(list);
Configuration conf = HBaseConfiguration.create();
JavaHBaseContext hbaseContext = new JavaHBaseContext(jsc, conf);
hbaseContext.foreachPartition(rdd,
new VoidFunction<Tuple2<Iterator<byte[]>, Connection>>() {
public void call(Tuple2<Iterator<byte[]>, Connection> t)
throws Exception {
Table table = t._2().getTable(TableName.valueOf(tableName));
BufferedMutator mutator = t._2().getBufferedMutator(TableName.valueOf(tableName));
while (t._1().hasNext()) {
byte[] b = t._1().next();
Result r = table.get(new Get(b));
if (r.getExists()) {
mutator.mutate(new Put(b));
}
}
mutator.flush();
mutator.close();
table.close();
}
});
} finally {
jsc.stop();
}
All functionality between Spark and HBase will be supported both in Scala and in Java, with the exception of SparkSQL which will support any language that is supported by Spark. For the remaining of this documentation we will focus on Scala examples for now.
The examples above illustrate how to do a foreachPartition with a connection. A number of other Spark base functions are supported out of the box:
bulkPut
-
For massively parallel sending of puts to HBase
bulkDelete
-
For massively parallel sending of deletes to HBase
bulkGet
-
For massively parallel sending of gets to HBase to create a new RDD
mapPartition
-
To do a Spark Map function with a Connection object to allow full access to HBase
hBaseRDD
-
To simplify a distributed scan to create a RDD
For examples of all these functionalities, see the HBase-Spark Module.
2. Spark Streaming
Spark Streaming is a micro batching stream processing framework built on top of Spark. HBase and Spark Streaming make great companions in that HBase can help serve the following benefits alongside Spark Streaming.
-
A place to grab reference data or profile data on the fly
-
A place to store counts or aggregates in a way that supports Spark Streaming promise of only once processing.
The HBase-Spark module’s integration points with Spark Streaming are similar to its normal Spark integration points, in that the following commands are possible straight off a Spark Streaming DStream.
bulkPut
-
For massively parallel sending of puts to HBase
bulkDelete
-
For massively parallel sending of deletes to HBase
bulkGet
-
For massively parallel sending of gets to HBase to create a new RDD
mapPartition
-
To do a Spark Map function with a Connection object to allow full access to HBase
hBaseRDD
-
To simplify a distributed scan to create a RDD
bulkPut
Example with DStreamsBelow is an example of bulkPut with DStreams. It is very close in feel to the RDD bulk put.
val sc = new SparkContext("local", "test")
val config = new HBaseConfiguration()
val hbaseContext = new HBaseContext(sc, config)
val ssc = new StreamingContext(sc, Milliseconds(200))
val rdd1 = ...
val rdd2 = ...
val queue = mutable.Queue[RDD[(Array[Byte], Array[(Array[Byte],
Array[Byte], Array[Byte])])]]()
queue += rdd1
queue += rdd2
val dStream = ssc.queueStream(queue)
dStream.hbaseBulkPut(
hbaseContext,
TableName.valueOf(tableName),
(putRecord) => {
val put = new Put(putRecord._1)
putRecord._2.foreach((putValue) => put.addColumn(putValue._1, putValue._2, putValue._3))
put
})
There are three inputs to the hbaseBulkPut
function.
. The hbaseContext that carries the configuration boardcast information link us
to the HBase Connections in the executors
. The table name of the table we are putting data into
. A function that will convert a record in the DStream into a HBase Put object.
3. Bulk Load
Spark bulk load follows very closely to the MapReduce implementation of bulk
load. In short, a partitioner partitions based on region splits and
the row keys are sent to the reducers in order, so that HFiles can be written
out. In Spark terms, the bulk load will be focused around a
repartitionAndSortWithinPartitions
followed by a foreachPartition
.
The only major difference with the Spark implementation compared to the MapReduce implementation is that the column qualifier is included in the shuffle ordering process. This was done because the MapReduce bulk load implementation would have memory issues with loading rows with a large numbers of columns, as a result of the sorting of those columns being done in the memory of the reducer JVM. Instead, that ordering is done in the Spark Shuffle, so there should no longer be a limit to the number of columns in a row for bulk loading.
The following example shows bulk loading with Spark.
val sc = new SparkContext("local", "test")
val config = new HBaseConfiguration()
val hbaseContext = new HBaseContext(sc, config)
val stagingFolder = ...
rdd.hbaseBulkLoad(TableName.valueOf(tableName),
t => {
val rowKey = t._1
val family:Array[Byte] = t._2(0)._1
val qualifier = t._2(0)._2
val value = t._2(0)._3
val keyFamilyQualifier= new KeyFamilyQualifier(rowKey, family, qualifier)
Seq((keyFamilyQualifier, value)).iterator
},
stagingFolder.getPath)
val load = new LoadIncrementalHFiles(config)
load.doBulkLoad(new Path(stagingFolder.getPath),
conn.getAdmin, table, conn.getRegionLocator(TableName.valueOf(tableName)))
The hbaseBulkLoad
function takes three required parameters:
-
The table name of the table we intend to bulk load too
-
A function that will convert a record in the RDD to a tuple key value par. With the tuple key being a KeyFamilyQualifer object and the value being the cell value. The KeyFamilyQualifer object will hold the RowKey, Column Family, and Column Qualifier. The shuffle will partition on the RowKey but will sort by all three values.
-
The temporary path for the HFile to be written out too
Following the Spark bulk load command, use the HBase’s LoadIncrementalHFiles object to load the newly created HFiles into HBase.
You can set the following attributes with additional parameter options on hbaseBulkLoad.
-
Max file size of the HFiles
-
A flag to exclude HFiles from compactions
-
Column Family settings for compression, bloomType, blockSize, and dataBlockEncoding
val sc = new SparkContext("local", "test")
val config = new HBaseConfiguration()
val hbaseContext = new HBaseContext(sc, config)
val stagingFolder = ...
val familyHBaseWriterOptions = new java.util.HashMap[Array[Byte], FamilyHFileWriteOptions]
val f1Options = new FamilyHFileWriteOptions("GZ", "ROW", 128, "PREFIX")
familyHBaseWriterOptions.put(Bytes.toBytes("columnFamily1"), f1Options)
rdd.hbaseBulkLoad(TableName.valueOf(tableName),
t => {
val rowKey = t._1
val family:Array[Byte] = t._2(0)._1
val qualifier = t._2(0)._2
val value = t._2(0)._3
val keyFamilyQualifier= new KeyFamilyQualifier(rowKey, family, qualifier)
Seq((keyFamilyQualifier, value)).iterator
},
stagingFolder.getPath,
familyHBaseWriterOptions,
compactionExclude = false,
HConstants.DEFAULT_MAX_FILE_SIZE)
val load = new LoadIncrementalHFiles(config)
load.doBulkLoad(new Path(stagingFolder.getPath),
conn.getAdmin, table, conn.getRegionLocator(TableName.valueOf(tableName)))
4. SparkSQL/DataFrames
SparkSQL is a subproject of Spark that supports SQL that will compute down to a Spark DAG. In addition,SparkSQL is a heavy user of DataFrames. DataFrames are like RDDs with schema information.
The HBase-Spark module includes support for Spark SQL and DataFrames, which allows you to write SparkSQL directly on HBase tables. In addition the HBase-Spark will push down query filtering logic to HBase.
4.1. Predicate Push Down
There are two examples of predicate push down in the HBase-Spark implementation. The first example shows the push down of filtering logic on the RowKey. HBase-Spark will reduce the filters on RowKeys down to a set of Get and/or Scan commands.
The Scans are distributed scans, rather than a single client scan operation. |
If the query looks something like the following, the logic will push down and get
the rows through 3 Gets and 0 Scans. We can do gets because all the operations
are equal
operations.
SELECT
KEY_FIELD,
B_FIELD,
A_FIELD
FROM hbaseTmp
WHERE (KEY_FIELD = 'get1' or KEY_FIELD = 'get2' or KEY_FIELD = 'get3')
Now lets look at an example where we will end up doing two scans on HBase.
SELECT
KEY_FIELD,
B_FIELD,
A_FIELD
FROM hbaseTmp
WHERE KEY_FIELD < 'get2' or KEY_FIELD > 'get3'
In this example we will get 0 Gets and 2 Scans. One scan will load everything from the first row in the table until “get2” and the second scan will get everything from “get3” until the last row in the table.
The next query is a good example of having a good deal of range checks. However the ranges overlap. To the code will be smart enough to get the following data in a single scan that encompasses all the data asked by the query.
SELECT
KEY_FIELD,
B_FIELD,
A_FIELD
FROM hbaseTmp
WHERE
(KEY_FIELD >= 'get1' and KEY_FIELD <= 'get3') or
(KEY_FIELD > 'get3' and KEY_FIELD <= 'get5')
The second example of push down functionality offered by the HBase-Spark module is the ability to push down filter logic for column and cell fields. Just like the RowKey logic, all query logic will be consolidated into the minimum number of range checks and equal checks by sending a Filter object along with the Scan with information about consolidated push down predicates
This example shows how we can interact with HBase with SQL.
val sc = new SparkContext("local", "test")
val config = new HBaseConfiguration()
new HBaseContext(sc, TEST_UTIL.getConfiguration)
val sqlContext = new SQLContext(sc)
df = sqlContext.load("org.apache.hadoop.hbase.spark",
Map("hbase.columns.mapping" ->
"KEY_FIELD STRING :key, A_FIELD STRING c:a, B_FIELD STRING c:b",
"hbase.table" -> "t1"))
df.registerTempTable("hbaseTmp")
val results = sqlContext.sql("SELECT KEY_FIELD, B_FIELD FROM hbaseTmp " +
"WHERE " +
"(KEY_FIELD = 'get1' and B_FIELD < '3') or " +
"(KEY_FIELD >= 'get3' and B_FIELD = '8')").take(5)
There are three major parts of this example that deserve explaining.
- The sqlContext.load function
-
In the sqlContext.load function we see two parameters. The first of these parameters is pointing Spark to the HBase DefaultSource class that will act as the interface between SparkSQL and HBase.
- A map of key value pairs
-
In this example we have two keys in our map,
hbase.columns.mapping
andhbase.table
. Thehbase.table
directs SparkSQL to use the given HBase table. Thehbase.columns.mapping
key give us the logic to translate HBase columns to SparkSQL columns.The
hbase.columns.mapping
is a string that follows the following format(SparkSQL.ColumnName) (SparkSQL.ColumnType) (HBase.ColumnFamily):(HBase.Qualifier)
In the example below we see the definition of three fields. Because KEY_FIELD has no ColumnFamily, it is the RowKey.
KEY_FIELD STRING :key, A_FIELD STRING c:a, B_FIELD STRING c:b
- The registerTempTable function
-
This is a SparkSQL function that allows us now to be free of Scala when accessing our HBase table directly with SQL with the table name of "hbaseTmp".
The last major point to note in the example is the sqlContext.sql
function, which
allows the user to ask their questions in SQL which will be pushed down to the
DefaultSource code in the HBase-Spark module. The result of this command will be
a DataFrame with the Schema of KEY_FIELD and B_FIELD.