简介
Apache Drill是一个低延迟的分布式海量数据（涵盖结构化、半结构化以及嵌套数据）交互式查询引擎，使用ANSI SQL兼容语法，支持本地文件、HDFS、HBase、MongoDB等后端存储，支持Parquet、JSON、CSV、TSV、PSV等数据格式。受Google的Dremel启发，Drill满足上千节点的PB级别数据的交互式商业智能分析场景。
安装
Drill可以安装在单机或者集群环境上，支持Linux、Windows、Mac OS X系统。简单起见，我们在Linux单机环境（CentOS 6.3）搭建以供试用。
准备安装包：
- jdk 7：jdk-7u75-linux-x64.tar.gz
- Drill：apache-drill-0.8.0.tar.gz
在$WORK（/path/to/work）目录中安装，将jdk和drill分别解压到java和drill目录中，并打软连以便升级：
.
├── drill
│ ├── apache-drill -> apache-drill-0.8.0
│ └── apache-drill-0.8.0
├── init.sh
└── java
├── jdk -> jdk1.7.0_75
└── jdk1.7.0_75
并添加一init.sh脚本初始化java相关环境变量：
export WORK="/path/to/work"
export JAVA="$WORK/java/jdk/bin/java"
export JAVA_HOME="$WORK/java/jdk"
启动
在单机环境运行只需要启动bin/sqlline便可：
$ cd $WORK
$ . ./init.sh
$ ./drill/apache-drill/bin/sqlline -u jdbc:drill:zk=local
Drill log directory /var/log/drill does not exist or is not writable, defaulting to ...
Apr 06, 2015 12:47:30 AM org.glassfish.jersey.server.ApplicationHandler initialize
INFO: Initiating Jersey application, version Jersey: 2.8 2014-04-29 01:25:26...
sqlline version 1.1.6
0: jdbc:drill:zk=local>
-u jdbc:drill:zk=local表示使用本机的Drill，无需启动ZooKeeper，如果是集群环境则需要配置和启动ZooKeeper并填写地址。启动后便可以在0: jdbc:drill:zk=local>后敲入命令使用了。
试用
Drill的sample-data目录有Parquet格式的演示数据可供查询：
0: jdbc:drill:zk=local> select * from dfs.`/path/to/work/drill/apache-drill/sample-data/nation.parquet` limit 5;
+-------------+------------+-------------+------------+
| N_NATIONKEY | N_NAME | N_REGIONKEY | N_COMMENT |
+-------------+------------+-------------+------------+
| 0 | ALGERIA | 0 | haggle. carefully f |
| 1 | ARGENTINA | 1 | al foxes promise sly |
| 2 | BRAZIL | 1 | y alongside of the p |
| 3 | CANADA | 1 | eas hang ironic, sil |
| 4 | EGYPT | 4 | y above the carefull |
+-------------+------------+-------------+------------+
5 rows selected (0.741 seconds)
这里用的库名格式为dfs.本地文件（Parquet、JSON、CSV等文件）绝对路径。可以看出只要熟悉SQL语法几乎没有学习成本。但Parquet格式文件需要专用工具查看、编辑，不是很方便，后续再专门介绍，下文先使用更通用的CSV和JSON文件进行演示。
在$WORK/data中创建如下test.csv文件：
1101,SteveEurich,Steve,Eurich,16,StoreT
1102,MaryPierson,Mary,Pierson,16,StoreT
1103,LeoJones,Leo,Jones,16,StoreTem
1104,NancyBeatty,Nancy,Beatty,16,StoreT
1105,ClaraMcNight,Clara,McNight,16,Store
然后查询：
0: jdbc:drill:zk=local> select * from dfs.`/path/to/work/drill/data/test.csv`;
+------------+
| columns |
+------------+
| ["1101","SteveEurich","Steve","Eurich","16","StoreT"] |
| ["1102","MaryPierson","Mary","Pierson","16","StoreT"] |
| ["1103","LeoJones","Leo","Jones","16","StoreTem"] |
| ["1104","NancyBeatty","Nancy","Beatty","16","StoreT"] |
| ["1105","ClaraMcNight","Clara","McNight","16","Store"] |
+------------+
5 rows selected (0.082 seconds)
可以看到结果和之前的稍有不同，因为CSV文件没有地方存放列列名，所以统一用columns代替，如果需要具体制定列则需要用columns[n]，如：
0: jdbc:drill:zk=local> select columns[0], columns[3] from dfs.`/path/to/work/drill/data/test.csv`;
+------------+------------+
| EXPR$0 | EXPR$1 |
+------------+------------+
| 1101 | Eurich |
| 1102 | Pierson |
| 1103 | Jones |
| 1104 | Beatty |
| 1105 | McNight |
+------------+------------+
CSV文件格式比较简单，发挥不出Drill的强大优势，下边更复杂的功能使用和Parquet更接近的JSON文件进行演示。
在$WORK/data中创建如下test.json文件：
{
"ka1": 1,
"kb1": 1.1,
"kc1": "vc11",
"kd1": [
{
"ka2": 10,
"kb2": 10.1,
"kc2": "vc1010"
}
]
}
{
"ka1": 2,
"kb1": 2.2,
"kc1": "vc22",
"kd1": [
{
"ka2": 20,
"kb2": 20.2,
"kc2": "vc2020"
}
]
}
{
"ka1": 3,
"kb1": 3.3,
"kc1": "vc33",
"kd1": [
{
"ka2": 30,
"kb2": 30.3,
"kc2": "vc3030"
}
]
}
可以看到这个JSON文件内容是有多层嵌套的，结构比之前那个CSV文件要复杂不少，而查询嵌套数据正是Drill的优势所在。
0: jdbc:drill:zk=local> select * from dfs.`/path/to/work/drill/data/test.json`;
+------------+------------+------------+------------+
| ka1 | kb1 | kc1 | kd1 |
+------------+------------+------------+------------+
| 1 | 1.1 | vc11 | [{"ka2":10,"kb2":10.1,"kc2":"vc1010"}] |
| 2 | 2.2 | vc22 | [{"ka2":20,"kb2":20.2,"kc2":"vc2020"}] |
| 3 | 3.3 | vc33 | [{"ka2":30,"kb2":30.3,"kc2":"vc3030"}] |
+------------+------------+------------+------------+
3 rows selected (0.098 seconds)
select *只查出第一层的数据，更深层的数据只以原本的JSON数据呈现出来，我们显然不应该只关心第一层的数据，具体怎么查完全随心所欲：
0: jdbc:drill:zk=local> select sum(ka1), avg(kd1[0].kb2) from dfs.`/path/to/work/drill/data/test.json`;
+------------+------------+
| EXPR$0 | EXPR$1 |
+------------+------------+
| 6 | 20.2 |
+------------+------------+
1 row selected (0.136 seconds)
可以通过kd1[0]来访问嵌套到第二层的这个表。
0: jdbc:drill:zk=local> select kc1, kd1[0].kc2 from dfs.`/path/to/work/drill/data/test.json` where kd1[0].kb2 = 10.1 and ka1 = 1;
+------------+------------+
| kc1 | EXPR$1 |
+------------+------------+
| vc11 | vc1010 |
+------------+------------+
1 row selected (0.181 seconds)
创建view：
0: jdbc:drill:zk=local> create view dfs.tmp.tmpview as select kd1[0].kb2 from dfs.`/path/to/work/drill/data/test.json`;
+------------+------------+
| ok | summary |
+------------+------------+
| true | View 'tmpview' created successfully in 'dfs.tmp' schema |
+------------+------------+
1 row selected (0.055 seconds)
0: jdbc:drill:zk=local> select * from dfs.tmp.tmpview;
+------------+
| EXPR$0 |
+------------+
| 10.1 |
| 20.2 |
| 30.3 |
+------------+
3 rows selected (0.193 seconds)
可以把嵌套的第二层表打平（整合kd1[0]..kd1[n]）：
0: jdbc:drill:zk=local> select kddb.kdtable.kc2 from (select flatten(kd1) kdtable from dfs.`/path/to/work/drill/data/test.json`) kddb;
+------------+
| EXPR$0 |
+------------+
| vc1010 |
| vc2020 |
| vc3030 |
+------------+
3 rows selected (0.083 seconds)
使用细节上和mysql还是有所不同的，另外涉及到多层表的复杂逻辑，要想用得得心应手还需要仔细阅读官方文档并多多练习。这次先走马观花了，之后会深入了解语法层面的特性。
网友评论