1、软件环境

1.1 大数据组件环境

大数据组件	版本
Hive	3.1.2
Spark	spark-3.0.0-bin-hadoop3.2

1.2 操作系统环境

OS	版本
MacOS	Monterey 12.1
Linux - CentOS	7.6

2、大数据组件搭建

2.1 Hive环境搭建

1）Hive on Spark说明

Hive引擎包括：默认 mr、spark、Tez。
Hive on Spark：Hive既作为存储元数据又负责SQL的解析优化，语法是HQL语法，执行引擎变成了Spark，Spark负责采用RDD执行。
Spark on Hive : Hive只作为存储元数据，Spark负责SQL解析优化，语法是Spark SQL语法，Spark负责采用RDD执行。

2）Hive on Spark配置

（1）兼容性说明

注意：官网下载的Hive3.1.2和Spark3.0.0默认是不兼容的。因为Hive3.1.2支持的Spark版本是2.4.5，所以需要我们重新编译Hive3.1.2版本。

编译步骤：官网下载Hive3.1.2源码，修改pom文件中引用的Spark版本为3.0.0，如果编译通过，直接打包获取jar包。如果报错，就根据提示，修改相关方法，直到不报错，打包获取jar包。

（2）在Hive所在节点部署Spark

如果之前已经部署了Spark，则该步骤可以跳过。

Spark官网下载jar包地址
http://spark.apache.org/downloads.html

上传并解压解压spark-3.0.0-bin-hadoop3.2.tgz
[postman@cdh01 software]$ tar -zxvf spark-3.0.0-bin-hadoop3.2.tgz -C /opt/module/

[postman@cdh01 software]$ mv /opt/module/spark-3.0.0-bin-hadoop3.2 /opt/module/spark

（3）配置SPARK_HOME环境变量

[postman@cdh01 software]$ sudo vim /etc/profile.d/my_env.sh
添加如下内容。

# SPARK_HOME
export SPARK_HOME=/opt/module/spark
export PATH=$PATH:$SPARK_HOME/bin

使其生效：
source ${配置脚本}

# For MacOS
[postman@cdh01 software]$ source ~/.zshrc

# For CentOS
[postman@cdh01 software]$ source /etc/profile.d/my_env.sh

（4）在hive中创建spark配置文件

[postman@cdh01 software]$ vim /opt/module/hive/conf/spark-defaults.conf

添加如下内容（在执行任务时，会根据如下参数执行）。

spark.master                             yarn
spark.eventLog.enabled                   true
spark.eventLog.dir                       hdfs://cdh01:8020/spark-history
spark.executor.memory                    1g
spark.driver.memory					     1g

在HDFS创建如下路径，用于存储历史日志。

[postman@cdh01 software]$ hadoop fs -mkdir /spark-history

（5）向HDFS上传Spark无 hadoop+hive 依赖的纯净jar包

• 说明1：由于Spark3.0.0非纯净版默认支持的是hive2.3.7版本，直接使用会和安装的Hive3.1.2出现兼容性问题。所以采用Spark纯净版jar包，不包含hadoop和hive相关依赖，避免冲突。
• 说明2：Hive任务最终由Spark来执行，Spark任务资源分配由Yarn来调度，该任务有可能被分配到集群的任何一个节点。所以需要将Spark的依赖上传到HDFS集群路径，这样集群中任何一个节点都能获取到。

上传并解压spark-3.0.0-bin-without-hadoop.tgz

[postman@cdh01 software]$ tar -zxf /opt/software/spark-3.0.0-bin-without-hadoop.tgz

上传Spark纯净版jar包到HDFS

[postman@cdh01 software]$ hadoop fs -mkdir -p /spark-jars
[postman@cdh01 software]$ hadoop fs -put spark-3.0.0-bin-without-hadoop/jars/* /spark-jars

（6）修改hive-site.xml文件

[postman@cdh01 ~]$ vim /opt/module/hive/conf/hive-site.xml

添加如下内容。

<!--Spark依赖位置（注意：端口号8020必须和namenode的端口号一致）-->
<property>
    <name>spark.yarn.jars</name>
    <value>hdfs://cdh01:8020/spark-jars/*</value>
</property>
  
<!--Hive执行引擎-->
<property>
    <name>hive.execution.engine</name>
    <value>spark</value>
</property>

7）修改 $SPARK_HOME/conf/spark-env.sh 文件

[postman@cdh01 ~]$ vim $SPARK_HOME/conf/spark-env.sh

添加如下内容。

export SPARK_DIST_CLASSPATH=$(hadoop classpath)

否则将报各类hadoop依赖包缺失的异常，如log4j、Hadoop的Configuration等包缺失。

2.2 Hive on Spark测试

（1）启动hive客户端

[postman@cdh01 hive]$ bin/hive

（2）创建一张测试表

hive (default)> create table user(id int, name string);

（3）通过insert测试效果

hive (default)> insert into table user values(1001,'zhangsan');

若结果如下，则说明配置成功。

hive (default)> insert into table user values(1001,'zhangsan');
Query ID = user_20231108165919_9908b655-96a7-4ccb-bb62-4dde28df9394
Total jobs = 1
Launching Job 1 out of 1
In order to change the average load for a reducer (in bytes):
  set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
  set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
  set mapreduce.job.reduces=<number>
Running with YARN Application = application_1699425455296_0013
Kill Command = /opt/module/hadoop-3.1.3/bin/yarn application -kill application_1699425455296_0013
Hive on Spark Session Web UI URL: http://192.168.1.1:60145

Query Hive on Spark job[0] stages: [0, 1]
Spark job[0] status = RUNNING
Job Progress Format
CurrentTime StageId_StageAttemptId: SucceededTasksCount(+RunningTasksCount-FailedTasksCount)/TotalTasksCount
2023-11-08 16:59:35,314 Stage-0_0: 0/1  Stage-1_0: 0/1
2023-11-08 16:59:37,331 Stage-0_0: 1/1 Finished Stage-1_0: 0/1
2023-11-08 16:59:39,363 Stage-0_0: 1/1 Finished Stage-1_0: 1/1 Finished
Spark job[0] finished successfully in 6.09 second(s)
Loading data to table default.user
OK
col1    col2
Time taken: 20.569 seconds
hive (default)> select * from user;
OK
user.id      user.name
1001         zhangsan

3、安装过程中的错误

3.1 M1芯片下 zstd 库文件错误

当执行 MR 类 sql 如 ”insert into table user values(1001,‘zhangsan’); “ 时，程序在 Console 上长时间卡住，但无错误日志输出，此时日志格式为：

hive (default)> insert into table student values(1,'abc');
Query ID = davidliu_20231108163620_eb8fabe4-b615-4d12-9dba-56ead5946a98
Total jobs = 1
Launching Job 1 out of 1
In order to change the average load for a reducer (in bytes):
  set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
  set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
  set mapreduce.job.reduces=<number>
Running with YARN Application = application_1699425455296_0010
Kill Command = /opt/module/hadoop-3.1.3/bin/yarn application -kill application_1699425455296_0010
Hive on Spark Session Web UI URL: http://192.168.154.240:56101

Query Hive on Spark job[0] stages: [0, 1]
Spark job[0] status = RUNNING
Job Progress Format
CurrentTime StageId_StageAttemptId: SucceededTasksCount(+RunningTasksCount-FailedTasksCount)/TotalTasksCount
2023-11-08 16:36:36,031 Stage-0_0: 0/1  Stage-1_0: 0/1
2023-11-08 16:36:39,089 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:36:42,148 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:36:45,201 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:36:48,270 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:36:51,331 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:36:54,385 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:36:57,435 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:00,478 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:03,517 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:06,572 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:09,606 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:12,653 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:15,700 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:18,737 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:21,790 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:24,832 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:27,874 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
2023-11-08 16:37:30,914 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
...
...
2023-11-08 16:37:33,974 Stage-0_0: 1/1 Finished Stage-1_0: 0(+1)/1
Interrupting... Be patient, this might take some time.
Press Ctrl+C again to kill JVM
Exiting the JVM

在“Ctrl + C”取消 sql 执行之前，去 yarn 控制页面查看了一下程序运行的结果：

在WebUI 页面上从某次失败 Application 的某次 MR 任务的执行 log 中，发现有如下错误：

Caused by: java.lang.UnsatisfiedLinkError: no zstd-jni in java.library.path
Unsupported OS/arch, cannot find /darwin/aarch64/libzstd-jni.dylib or load zstd-jni from system libraries. Please try building from source the jar or providing libzstd-jni in your system.
       at java.lang.Runtime.loadLibrary0(Runtime.java:1011)
       at java.lang.System.loadLibrary(System.java:1657)
       at com.github.luben.zstd.util.Native.load(Native.java:85)
       at com.github.luben.zstd.util.Native.load(Native.java:55)
       at com.github.luben.zstd.Zstd.<clinit>(Zstd.java:13)
       at com.github.luben.zstd.Zstd.decompressedSize(Zstd.java:579)

同时在 Hadoop ResourceManager 的运行日志中也发现了关于这块的报错日志。

从上述 log 中可以看出zstd 软件库包(作用：文件压缩)在 M1 芯片下 不能很高的被支持，结合 Hive On Spark 运行的库包路径查找比对，最终在上传到HDFS集群路径/spark-jars 下 Hive on Spark的依赖jar 包中发现了 zstd jar 包：

• zstd-jni-1.4.4-3.jar

经查，此前已有开发者在 zstd 的github项目 下上报过这个问题，且有网友反馈在"1.4.9-1"版本中已修复了该问题。
于是在 mvnrepository 网站 上下载版本的 jar 包：

• zstd-jni-1.4.9-1.jar

之后，将 HDFS 路径“hdfs://cdh01:8020/spark-jars/*”下的原始 “zstd-jni-1.4.4-3.jar” 删除，并替换为 “zstd-jni-1.4.9-1.jar” 后（如上图所示），经再度测试，该问题就解决了。