Apache Spark Plugin

The phoenix-spark plugin extends Phoenix’s MapReduce support to allow Spark to load Phoenix tables as RDDs or DataFrames, and enables persisting them back to Phoenix.

Prerequisites

  • Phoenix 4.4.0+
  • Spark 1.3.1+ (prebuilt with Hadoop 2.4 recommended)

Why not JDBC?

Although Spark supports connecting directly to JDBC databases, it’s only able to parallelize queries by partioning on a numeric column. It also requires a known lower bound, upper bound and partition count in order to create split queries.

In contrast, the phoenix-spark integration is able to leverage the underlying splits provided by Phoenix in order to retrieve and save data across multiple workers. All that’s required is a database URL and a table name. Optional SELECT columns can be given, as well as pushdown predicates for efficient filtering.

The choice of which method to use to access Phoenix comes down to each specific use case.

Spark setup

  • To ensure that all requisite Phoenix / HBase platform dependencies are available on the classpath for the Spark executors and drivers, set both ‘spark.executor.extraClassPath’ and ‘spark.driver.extraClassPath’ in spark-defaults.conf to include the ‘phoenix--client.jar’

  • Note that for Phoenix versions 4.7 and 4.8 you must use the ‘phoenix--client-spark.jar’. As of Phoenix 4.10, the ‘phoenix--client.jar’ is compiled against Spark 2.x. If compability with Spark 1.x if needed, you must compile Phoenix with the spark16 maven profile.

  • To help your IDE, you can add the following provided dependency to your build:

  1. <dependency>
  2.   <groupId>org.apache.phoenix</groupId>
  3.   <artifactId>phoenix-spark</artifactId>
  4.   <version>${phoenix.version}</version>
  5.   <scope>provided</scope>
  6. </dependency>

Reading Phoenix Tables

Given a Phoenix table with the following DDL

  1. CREATE TABLE TABLE1 (ID BIGINT NOT NULL PRIMARY KEY, COL1 VARCHAR);
  2. UPSERT INTO TABLE1 (ID, COL1) VALUES (1, 'test_row_1');
  3. UPSERT INTO TABLE1 (ID, COL1) VALUES (2, 'test_row_2');

Load as a DataFrame using the Data Source API

  1. import org.apache.spark.SparkContext
  2. import org.apache.spark.sql.SQLContext
  3. import org.apache.phoenix.spark._
  4.  
  5. val sc = new SparkContext("local", "phoenix-test")
  6. val sqlContext = new SQLContext(sc)
  7.  
  8. val df = sqlContext.load(
  9.   "org.apache.phoenix.spark",
  10.   Map("table" -> "TABLE1", "zkUrl" -> "phoenix-server:2181")
  11. )
  12.  
  13. df
  14.   .filter(df("COL1") === "test_row_1" && df("ID") === 1L)
  15.   .select(df("ID"))
  16.   .show

Load as a DataFrame directly using a Configuration object

  1. import org.apache.hadoop.conf.Configuration
  2. import org.apache.spark.SparkContext
  3. import org.apache.spark.sql.SQLContext
  4. import org.apache.phoenix.spark._
  5.  
  6. val configuration = new Configuration()
  7. // Can set Phoenix-specific settings, requires 'hbase.zookeeper.quorum'
  8.  
  9. val sc = new SparkContext("local", "phoenix-test")
  10. val sqlContext = new SQLContext(sc)
  11.  
  12. // Load the columns 'ID' and 'COL1' from TABLE1 as a DataFrame
  13. val df = sqlContext.phoenixTableAsDataFrame(
  14.   "TABLE1", Array("ID", "COL1"), conf = configuration
  15. )
  16.  
  17. df.show

Load as an RDD, using a Zookeeper URL

  1. import org.apache.spark.SparkContext
  2. import org.apache.spark.sql.SQLContext
  3. import org.apache.phoenix.spark._
  4.  
  5. val sc = new SparkContext("local", "phoenix-test")
  6.  
  7. // Load the columns 'ID' and 'COL1' from TABLE1 as an RDD
  8. val rdd: RDD[Map[String, AnyRef]] = sc.phoenixTableAsRDD(
  9.   "TABLE1", Seq("ID", "COL1"), zkUrl = Some("phoenix-server:2181")
  10. )
  11.  
  12. rdd.count()
  13.  
  14. val firstId = rdd1.first()("ID").asInstanceOf[Long]
  15. val firstCol = rdd1.first()("COL1").asInstanceOf[String]

Saving Phoenix

Given a Phoenix table with the following DDL

  1. CREATE TABLE OUTPUT_TEST_TABLE (id BIGINT NOT NULL PRIMARY KEY, col1 VARCHAR, col2 INTEGER);

Saving RDDs

The saveToPhoenix method is an implicit method on RDD[Product], or an RDD of Tuples. The data types must correspond to one of the Java types supported by Phoenix.

  1. import org.apache.spark.SparkContext
  2. import org.apache.phoenix.spark._
  3.  
  4. val sc = new SparkContext("local", "phoenix-test")
  5. val dataSet = List((1L, "1", 1), (2L, "2", 2), (3L, "3", 3))
  6.  
  7. sc
  8.   .parallelize(dataSet)
  9.   .saveToPhoenix(
  10.     "OUTPUT_TEST_TABLE",
  11.     Seq("ID","COL1","COL2"),
  12.     zkUrl = Some("phoenix-server:2181")
  13.   )

Saving DataFrames

The save is method on DataFrame allows passing in a data source type. You can use org.apache.phoenix.spark, and must also pass in a table and zkUrl parameter to specify which table and server to persist the DataFrame to. The column names are derived from the DataFrame’s schema field names, and must match the Phoenix column names.

The save method also takes a SaveMode option, for which only SaveMode.Overwrite is supported.

Given two Phoenix tables with the following DDL:

  1. CREATE TABLE INPUT_TABLE (id BIGINT NOT NULL PRIMARY KEY, col1 VARCHAR, col2 INTEGER);
  2. CREATE TABLE OUTPUT_TABLE (id BIGINT NOT NULL PRIMARY KEY, col1 VARCHAR, col2 INTEGER);
  1. import org.apache.spark.SparkContext
  2. import org.apache.spark.sql._
  3. import org.apache.phoenix.spark._
  4.  
  5. // Load INPUT_TABLE
  6. val sc = new SparkContext("local", "phoenix-test")
  7. val sqlContext = new SQLContext(sc)
  8. val df = sqlContext.load("org.apache.phoenix.spark", Map("table" -> "INPUT_TABLE",
  9.   "zkUrl" -> hbaseConnectionString))
  10.  
  11. // Save to OUTPUT_TABLE
  12. df.save("org.apache.phoenix.spark", SaveMode.Overwrite, Map("table" -> "OUTPUT_TABLE",
  13.   "zkUrl" -> hbaseConnectionString))

PySpark

With Spark’s DataFrame support, you can also use pyspark to read and write from Phoenix tables.

Load a DataFrame

Given a table TABLE1 and a Zookeeper url of localhost:2181 you can load the table as a DataFrame using the following Python code in pyspark

  1. df = sqlContext.read \
  2.   .format("org.apache.phoenix.spark") \
  3.   .option("table", "TABLE1") \
  4.   .option("zkUrl", "localhost:2181") \
  5.   .load()

Save a DataFrame

Given the same table and Zookeeper URLs above, you can save a DataFrame to a Phoenix table using the following code

  1. df.write \
  2.   .format("org.apache.phoenix.spark") \
  3.   .mode("overwrite") \
  4.   .option("table", "TABLE1") \
  5.   .option("zkUrl", "localhost:2181") \
  6.   .save()

Notes

The functions phoenixTableAsDataFrame, phoenixTableAsRDD and saveToPhoenix all support optionally specifying a conf Hadoop configuration parameter with custom Phoenix client settings, as well as an optional zkUrl parameter for the Phoenix connection URL.

If zkUrl isn’t specified, it’s assumed that the “hbase.zookeeper.quorum” property has been set in the conf parameter. Similarly, if no configuration is passed in, zkUrl must be specified.

Limitations

  • Basic support for column and predicate pushdown using the Data Source API
  • The Data Source API does not support passing custom Phoenix settings in configuration, you must create the DataFrame or RDD directly if you need fine-grained configuration.
  • No support for aggregate or distinct queries as explained in our Map Reduce Integration documentation.

PageRank example

This example makes use of the Enron email data set, provided by the Stanford Network Analysis Project, and executes the GraphX implementation of PageRank on it to find interesting entities. It then saves the results back to Phoenix.

  • Download and extract the file enron.csv.gz

  • Create the necessary Phoenix schema

CREATE TABLE EMAIL_ENRON(MAIL_FROM BIGINT NOT NULL, MAIL_TO BIGINT NOT NULL CONSTRAINT pk PRIMARY KEY(MAIL_FROM, MAIL_TO));
CREATE TABLE EMAIL_ENRON_PAGERANK(ID BIGINT NOT NULL, RANK DOUBLE CONSTRAINT pk PRIMARY KEY(ID));

  • Load the email data into Phoenix (assuming localhost for Zookeeper Quroum URL)

gunzip /tmp/enron.csv.gz
cd /path/to/phoenix/bin
./psql.py -t EMAIL_ENRON localhost /tmp/enron.csv

  • In spark-shell, with the phoenix-client in the Spark driver classpath, run the following:

import org.apache.spark.graphx.
import org.apache.phoenix.spark.
val rdd = sc.phoenixTableAsRDD("EMAIL_ENRON", Seq("MAIL_FROM", "MAIL_TO"), zkUrl=Some("localhost")) // load from phoenix
val rawEdges = rdd.map{ e => (e("MAIL_FROM").asInstanceOf[VertexId], e("MAIL_TO").asInstanceOf[VertexId]) } // map to vertexids
val graph = Graph.fromEdgeTuples(rawEdges, 1.0) // create a graph
val pr = graph.pageRank(0.001) // run pagerank
pr.vertices.saveToPhoenix("EMAIL_ENRON_PAGERANK", Seq("ID", "RANK"), zkUrl = Some("localhost")) // save to phoenix

  • Query the top ranked entities in SQL

SELECT * FROM EMAIL_ENRON_PAGERANK ORDER BY RANK DESC LIMIT 5;
+—————————————————————+—————————————————————+
| ID | RANK |
+—————————————————————+—————————————————————+
| 5038 | 497.2989872977676 |
| 273 | 117.18141799210386 |
| 140 | 108.63091596789913 |
| 458 | 107.2728800448782 |
| 588 | 106.11840798585399 |
+—————————————————————+—————————————————————+

原文: http://phoenix.apache.org/phoenix_spark.html