Spark SQL with Hive

前一篇文章是Spark SQL的入门篇Spark SQL初探，介绍了一些基础知识和API，可是离我们的日常使用还似乎差了一步之遥。

终结Shark的利用有2个：

1、和Spark程序的集成有诸多限制

2、Hive的优化器不是为Spark而设计的，计算模型的不同，使得Hive的优化器来优化Spark程序遇到了瓶颈。

这里看一下Spark SQL 的基础架构：

Spark1.1公布后会支持Spark SQL CLI 。 Spark SQL的CLI会要求被连接到一个Hive Thrift Server上，来实现类似hive shell的功能。

（ps：眼下git里面的branch-1.0-jdbc。眼下还没有正式release，我測了一下午。发现还是有bug的，耐心等待release吧！）

本着研究的心态，想和Hive环境集成一下。在spark shell里运行hive的语句。

一、编译Spark支持Hive

让Spark支持Hive有2种sbt编译方式：

1、sbt前加变量名

SPARK_HADOOP_VERSION=0.20.2-cdh3u5 SPARK_HIVE=true sbt/sbt assembly

2、改动project/SparkBuild.scala文件

val DEFAULT_HADOOP_VERSION = "0.20.2-cdh3u5"
val DEFAULT_HIVE = true 

然后运行sbt/sbt assembly

二、Spark SQL 操作Hive

前置：hive可用，而且在Spark-env.sh下，须要将Hive的conf和Hadoop的conf配到CLASSPATH里。

启动spark-shell

[root@web01 spark]# bin/spark-shell --master spark://10.1.8.210:7077 --driver-class-path /app/hadoop/hive-0.11.0-bin/lib/mysql-connector-java-5.1.13-bin.jar:/app/hadoop/hive-0.11.0-bin/lib/hadoop-lzo-0.4.15.jar

导入HiveContext

scala> val hiveContext = new org.apache.spark.sql.hive.HiveContext(sc)
hiveContext: org.apache.spark.sql.hive.HiveContext = org.apache.spark.sql.hive.HiveContext@7766d31c

scala> import hiveContext._
import hiveContext._

hiveContext里提供了一个运行sql的函数 hql(string text)

去hive里show databases. 这里Spark会parse hql 然后生成Query Plan。可是这里不会运行查询，仅仅有调用collect的时候才会运行。

scala> val show_databases = hql("show databases")
14/07/09 19:59:09 INFO storage.BlockManager: Removing broadcast 0
14/07/09 19:59:09 INFO storage.BlockManager: Removing block broadcast_0
14/07/09 19:59:09 INFO parse.ParseDriver: Parsing command: show databases
14/07/09 19:59:09 INFO parse.ParseDriver: Parse Completed
14/07/09 19:59:09 INFO analysis.Analyzer: Max iterations (2) reached for batch MultiInstanceRelations
14/07/09 19:59:09 INFO analysis.Analyzer: Max iterations (2) reached for batch CaseInsensitiveAttributeReferences
14/07/09 19:59:09 INFO analysis.Analyzer: Max iterations (2) reached for batch Check Analysis
14/07/09 19:59:09 INFO storage.MemoryStore: Block broadcast_0 of size 393044 dropped from memory (free 308713881)
14/07/09 19:59:09 INFO broadcast.HttpBroadcast: Deleted broadcast file: /tmp/spark-c29da0f8-c5e3-4fbf-adff-9aa77f9743b2/broadcast_0
14/07/09 19:59:09 INFO sql.SQLContext$$anon$1: Max iterations (2) reached for batch Add exchange
14/07/09 19:59:09 INFO sql.SQLContext$$anon$1: Max iterations (2) reached for batch Prepare Expressions
14/07/09 19:59:09 INFO spark.ContextCleaner: Cleaned broadcast 0
14/07/09 19:59:09 INFO ql.Driver: <PERFLOG method=Driver.run>
14/07/09 19:59:09 INFO ql.Driver: <PERFLOG method=TimeToSubmit>
14/07/09 19:59:09 INFO ql.Driver: <PERFLOG method=compile>
14/07/09 19:59:09 INFO exec.ListSinkOperator: 0 finished. closing...
14/07/09 19:59:09 INFO exec.ListSinkOperator: 0 forwarded 0 rows
14/07/09 19:59:09 INFO ql.Driver: <PERFLOG method=parse>
14/07/09 19:59:09 INFO parse.ParseDriver: Parsing command: show databases
14/07/09 19:59:09 INFO parse.ParseDriver: Parse Completed
14/07/09 19:59:09 INFO ql.Driver: </PERFLOG method=parse start=1404907149927 end=1404907149928 duration=1>
14/07/09 19:59:09 INFO ql.Driver: <PERFLOG method=semanticAnalyze>
14/07/09 19:59:09 INFO ql.Driver: Semantic Analysis Completed
14/07/09 19:59:09 INFO ql.Driver: </PERFLOG method=semanticAnalyze start=1404907149928 end=1404907149977 duration=49>
14/07/09 19:59:09 INFO exec.ListSinkOperator: Initializing Self 0 OP
14/07/09 19:59:09 INFO exec.ListSinkOperator: Operator 0 OP initialized
14/07/09 19:59:09 INFO exec.ListSinkOperator: Initialization Done 0 OP
14/07/09 19:59:09 INFO ql.Driver: Returning Hive schema: Schema(fieldSchemas:[FieldSchema(name:database_name, type:string, comment:from deserializer)], properties:null)
14/07/09 19:59:09 INFO ql.Driver: </PERFLOG method=compile start=1404907149925 end=1404907149980 duration=55>
14/07/09 19:59:09 INFO ql.Driver: <PERFLOG method=Driver.execute>
14/07/09 19:59:09 INFO ql.Driver: Starting command: show databases
14/07/09 19:59:09 INFO ql.Driver: </PERFLOG method=TimeToSubmit start=1404907149925 end=1404907149980 duration=55>
14/07/09 19:59:09 INFO ql.Driver: <PERFLOG method=runTasks>
14/07/09 19:59:09 INFO ql.Driver: <PERFLOG method=task.DDL.Stage-0>
14/07/09 19:59:09 INFO metastore.HiveMetaStore: 0: get_all_databases
14/07/09 19:59:09 INFO HiveMetaStore.audit: ugi=root    ip=unknown-ip-addr      cmd=get_all_databases
14/07/09 19:59:09 INFO exec.DDLTask: results : 1
14/07/09 19:59:10 INFO ql.Driver: </PERFLOG method=task.DDL.Stage-0 start=1404907149980 end=1404907150032 duration=52>
14/07/09 19:59:10 INFO ql.Driver: </PERFLOG method=runTasks start=1404907149980 end=1404907150032 duration=52>
14/07/09 19:59:10 INFO ql.Driver: </PERFLOG method=Driver.execute start=1404907149980 end=1404907150032 duration=52>
14/07/09 19:59:10 INFO ql.Driver: OK
14/07/09 19:59:10 INFO ql.Driver: <PERFLOG method=releaseLocks>
14/07/09 19:59:10 INFO ql.Driver: </PERFLOG method=releaseLocks start=1404907150033 end=1404907150033 duration=0>
14/07/09 19:59:10 INFO ql.Driver: </PERFLOG method=Driver.run start=1404907149925 end=1404907150033 duration=108>
14/07/09 19:59:10 INFO mapred.FileInputFormat: Total input paths to process : 1
14/07/09 19:59:10 INFO ql.Driver: <PERFLOG method=releaseLocks>
14/07/09 19:59:10 INFO ql.Driver: </PERFLOG method=releaseLocks start=1404907150037 end=1404907150037 duration=0>
show_databases: org.apache.spark.sql.SchemaRDD =
SchemaRDD[16] at RDD at SchemaRDD.scala:100
== Query Plan ==
<Native command: executed by Hive>

运行查询计划：

scala> show_databases.collect()
14/07/09 20:00:44 INFO spark.SparkContext: Starting job: collect at SparkPlan.scala:52
14/07/09 20:00:44 INFO scheduler.DAGScheduler: Got job 2 (collect at SparkPlan.scala:52) with 1 output partitions (allowLocal=false)
14/07/09 20:00:44 INFO scheduler.DAGScheduler: Final stage: Stage 2(collect at SparkPlan.scala:52)
14/07/09 20:00:44 INFO scheduler.DAGScheduler: Parents of final stage: List()
14/07/09 20:00:44 INFO scheduler.DAGScheduler: Missing parents: List()
14/07/09 20:00:44 INFO scheduler.DAGScheduler: Submitting Stage 2 (MappedRDD[20] at map at SparkPlan.scala:52), which has no missing parents
14/07/09 20:00:44 INFO scheduler.DAGScheduler: Submitting 1 missing tasks from Stage 2 (MappedRDD[20] at map at SparkPlan.scala:52)
14/07/09 20:00:44 INFO scheduler.TaskSchedulerImpl: Adding task set 2.0 with 1 tasks
14/07/09 20:00:44 INFO scheduler.TaskSetManager: Starting task 2.0:0 as TID 9 on executor 0: web01.dw (PROCESS_LOCAL)
14/07/09 20:00:44 INFO scheduler.TaskSetManager: Serialized task 2.0:0 as 1511 bytes in 0 ms
14/07/09 20:00:45 INFO scheduler.DAGScheduler: Completed ResultTask(2, 0)
14/07/09 20:00:45 INFO scheduler.TaskSetManager: Finished TID 9 in 12 ms on web01.dw (progress: 1/1)
14/07/09 20:00:45 INFO scheduler.TaskSchedulerImpl: Removed TaskSet 2.0, whose tasks have all completed, from pool
14/07/09 20:00:45 INFO scheduler.DAGScheduler: Stage 2 (collect at SparkPlan.scala:52) finished in 0.014 s
14/07/09 20:00:45 INFO spark.SparkContext: Job finished: collect at SparkPlan.scala:52, took 0.020520428 s
res5: Array[org.apache.spark.sql.Row] = Array([default])

返回default数据库。

相同的运行：show tables

scala> hql("show tables").collect()

14/07/09 20:01:28 INFO scheduler.TaskSchedulerImpl: Removed TaskSet 3.0, whose tasks have all completed, from pool
14/07/09 20:01:28 INFO scheduler.DAGScheduler: Stage 3 (collect at SparkPlan.scala:52) finished in 0.013 s
14/07/09 20:01:28 INFO spark.SparkContext: Job finished: collect at SparkPlan.scala:52, took 0.019173851 s
res7: Array[org.apache.spark.sql.Row] = Array([item], [src])

理论上是支持HIVE全部的操作，包含UDF。

PS:遇到的问题：

Caused by: org.datanucleus.exceptions.NucleusException: Attempt to invoke the "BoneCP" plugin to create a ConnectionPool gave an error : The specified datastore driver ("com.mysql.jdbc.Driver") was not found in the CLASSPATH. Please check your CLASSPATH specification, and the name of the driver.

解决的方法：就是我上面启动的时候带上sql-connector的路径。

。

三、总结：

Spark SQL 兼容了Hive的大部分语法和UDF，可是在处理查询计划的时候，使用了Catalyst框架进行优化，优化成适合Spark编程模型的运行计划，使得效率上高出hive非常多。

因为Spark1.1临时还未公布。眼下还存在bug，等到稳定版公布了再继续測试了。

全文完：）

原创文章，转载请注明出自：http://blog.csdn.net/oopsoom/article/details/37603261