摘要：本文由阿里巴巴的陈玉兆分享，主要介绍 Flink 集成 Hudi 的最新版本功能以及快速上手实践指南。内容包括：
- 背景
- 环境准备
- Batch 模式的读写
- Streaming 读
- 总结
一、背景
Apache Hudi 是目前最流行的数据湖解决方案之一，Data Lake Analytics[1] 集成了 Hudi 服务高效的数据 MERGE（UPDATE/DELETE）场景；AWS 在 EMR 服务中 预安装[2] 了 Apache Hudi，为用户提供高效的 record-level updates/deletes 和高效的数据查询管理；Uber [3]已经稳定运行 Apache Hudi 服务 4 年多，提供了低延迟的数据库同步和高效率的查询[4]。自 2016 年 8 月上线以来，数据湖存储规模已经超过 100PB[5]。
Apache Flink 作为目前最流行的流计算框架，在流式计算场景有天然的优势，当前，Flink 社区也在积极拥抱 Hudi 社区，发挥自身 streaming 写/读的优势，同时也对 batch 的读写做了支持。
Hudi 和 Fink 在 0.8.0 版本做了大量的集成工作[6]。核心的功能包括：
- 实现了新的 Flink streaming writer
- 支持 batch 和 streaming 模式 reader
- 支持 Flink SQL API
Flink streaming writer 通过 state 实现了高效的 index 方案，同时 Hudi 在 UPDATE/DELETE 上的优秀设计使得 Flink Hudi 成为当前最有潜力的 CDC 数据入湖方案，因为篇幅关系，将在后续的文章中介绍。
本文用 Flink SQL Client 来简单的演示通过 Flink SQL API 的方式实现 Hudi 表的操作，包括 batch 模式的读写和 streaming 模式的读。
二、环境准备
本文使用 Flink Sql Client[7] 作为演示工具，SQL CLI 可以比较方便地执行 SQL 的交互操作。
第一步：下载 Flink jar
Hudi 集成了 Flink 的 1.11 版本。您可以参考这里[8]来设置 Flink 环境。hudi-flink-bundle jar 是一个集成了 Flink 相关的 jar 的 uber jar, 目前推荐使用 scala 2.11 来编译。
第二步：设置 Flink 集群
启动一个 standalone 的 Flink 集群。启动之前，建议将 Flink 的集群配置设置如下：
- 在 $FLINK_HOME/conf/flink-conf.yaml 中添加配置项 taskmanager.numberOfTaskSlots: 4
- 在 $FLINK_HOME/conf/workers 中将条目 localhost 设置成 4 行，这里的行数代表了本地启动的 worker 数
启动集群：
# HADOOP_HOME is your hadoop root directory after unpack the binary package.
export HADOOP_CLASSPATH=`$HADOOP_HOME/bin/hadoop classpath`
# Start the flink standalone cluster
./bin/start-cluster.sh
第三步：启动 Flink SQL Client
Hudi 的 bundle jar 应该在 Sql Client 启动的时候加载到 CLASSPATH 中。您可以在路径 hudi-source-dir/packaging/hudi-flink-bundle 下手动编译 jar 包或者从 Apache Official Repository [9]下载。
启动 SQL CLI:
# HADOOP_HOME is your hadoop root directory after unpack the binary package.
export HADOOP_CLASSPATH=`$HADOOP_HOME/bin/hadoop classpath`
./bin/sql-client.sh embedded -j .../hudi-flink-bundle_2.1?-*.*.*.jar shell
备注：
- 推荐使用 hadoop 2.9.x+ 版本，因为一些对象存储（aliyun-oss）从这个版本开始支持
- flink-parquet 和 flink-avro 已经被打进 hudi-flink-bundle jar
- 您也可以直接将 hudi-flink-bundle jar 拷贝到 $FLINK_HOME/lib 目录下
- 本文的存储选取了对象存储 aliyun-oss，为了方便，您也可以使用本地路径
演示的工作目录结构如下：
/Users/chenyuzhao/workspace/hudi-demo
/- flink-1.11.3
/- hadoop-2.9.2
三、Batch 模式的读写
插入数据
使用如下 DDL 语句创建 Hudi 表：
Flink SQL> create table t2(
> uuid varchar(20),
> name varchar(10),
> age int,
> ts timestamp(3),
> `partition` varchar(20)
> )
> PARTITIONED BY (`partition`)
> with (
> 'connector' = 'hudi',
> 'path' = 'oss://vvr-daily/hudi/t2'
> );
[INFO] Table has been created.
DDL 里申明了表的 path，record key 为默认值 uuid，pre-combine key 为默认值 ts 。
然后通过 VALUES 语句往表中插入数据：
Flink SQL> insert into t2 values
> ('id1','Danny',23,TIMESTAMP '1970-01-01 00:00:01','par1'),
> ('id2','Stephen',33,TIMESTAMP '1970-01-01 00:00:02','par1'),
> ('id3','Julian',53,TIMESTAMP '1970-01-01 00:00:03','par2'),
> ('id4','Fabian',31,TIMESTAMP '1970-01-01 00:00:04','par2'),
> ('id5','Sophia',18,TIMESTAMP '1970-01-01 00:00:05','par3'),
> ('id6','Emma',20,TIMESTAMP '1970-01-01 00:00:06','par3'),
> ('id7','Bob',44,TIMESTAMP '1970-01-01 00:00:07','par4'),
> ('id8','Han',56,TIMESTAMP '1970-01-01 00:00:08','par4');
[INFO] Submitting SQL update statement to the cluster...
[INFO] Table update statement has been successfully submitted to the cluster:
Job ID: 59f2e528d14061f23c552a7ebf9a76bd
这里看到 Flink 的作业已经成功提交到集群，可以本地打开 web UI 观察作业的执行情况：
查询数据
作业执行完成后，通过 SELECT 语句查询表结果：
Flink SQL> set execution.result-mode=tableau;
[INFO] Session property has been set.
Flink SQL> select * from t2;
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| +/- | uuid | name | age | ts | partition |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| + | id3 | Julian | 53 | 1970-01-01T00:00:03 | par2 |
| + | id4 | Fabian | 31 | 1970-01-01T00:00:04 | par2 |
| + | id7 | Bob | 44 | 1970-01-01T00:00:07 | par4 |
| + | id8 | Han | 56 | 1970-01-01T00:00:08 | par4 |
| + | id1 | Danny | 23 | 1970-01-01T00:00:01 | par1 |
| + | id2 | Stephen | 33 | 1970-01-01T00:00:02 | par1 |
| + | id5 | Sophia | 18 | 1970-01-01T00:00:05 | par3 |
| + | id6 | Emma | 20 | 1970-01-01T00:00:06 | par3 |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
Received a total of 8 rows
这里执行语句 set execution.result-mode=tableau; 可以让查询结果直接输出到终端。
通过在 WHERE 子句中添加 partition 路径来裁剪 partition：
Flink SQL> select * from t2 where `partition` = 'par1';
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| +/- | uuid | name | age | ts | partition |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| + | id1 | Danny | 23 | 1970-01-01T00:00:01 | par1 |
| + | id2 | Stephen | 33 | 1970-01-01T00:00:02 | par1 |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
Received a total of 2 rows
更新数据
相同的 record key 的数据会自动覆盖，通过 INSERT 相同 key 的数据可以实现数据更新：
Flink SQL> insert into t2 values
> ('id1','Danny',24,TIMESTAMP '1970-01-01 00:00:01','par1'),
> ('id2','Stephen',34,TIMESTAMP '1970-01-01 00:00:02','par1');
[INFO] Submitting SQL update statement to the cluster...
[INFO] Table update statement has been successfully submitted to the cluster:
Job ID: 944de5a1ecbb7eeb4d1e9e748174fe4c
Flink SQL> select * from t2;
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| +/- | uuid | name | age | ts | partition |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| + | id1 | Danny | 24 | 1970-01-01T00:00:01 | par1 |
| + | id2 | Stephen | 34 | 1970-01-01T00:00:02 | par1 |
| + | id3 | Julian | 53 | 1970-01-01T00:00:03 | par2 |
| + | id4 | Fabian | 31 | 1970-01-01T00:00:04 | par2 |
| + | id5 | Sophia | 18 | 1970-01-01T00:00:05 | par3 |
| + | id6 | Emma | 20 | 1970-01-01T00:00:06 | par3 |
| + | id7 | Bob | 44 | 1970-01-01T00:00:07 | par4 |
| + | id8 | Han | 56 | 1970-01-01T00:00:08 | par4 |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
Received a total of 8 rows
可以看到 uuid 为 id1 和 id2 的数据 age 字段值发生了更新。
再次 insert 新数据观察结果：
Flink SQL> insert into t2 values
> ('id4','Fabian',32,TIMESTAMP '1970-01-01 00:00:04','par2'),
> ('id5','Sophia',19,TIMESTAMP '1970-01-01 00:00:05','par3');
[INFO] Submitting SQL update statement to the cluster...
[INFO] Table update statement has been successfully submitted to the cluster:
Job ID: fdeb7fd9f08808e66d77220f43075720
Flink SQL> select * from t2;
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| +/- | uuid | name | age | ts | partition |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| + | id5 | Sophia | 19 | 1970-01-01T00:00:05 | par3 |
| + | id6 | Emma | 20 | 1970-01-01T00:00:06 | par3 |
| + | id3 | Julian | 53 | 1970-01-01T00:00:03 | par2 |
| + | id4 | Fabian | 32 | 1970-01-01T00:00:04 | par2 |
| + | id1 | Danny | 24 | 1970-01-01T00:00:01 | par1 |
| + | id2 | Stephen | 34 | 1970-01-01T00:00:02 | par1 |
| + | id7 | Bob | 44 | 1970-01-01T00:00:07 | par4 |
| + | id8 | Han | 56 | 1970-01-01T00:00:08 | par4 |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
Received a total of 8 rows
四、Streaming 读
通过如下语句创建一张新的表并注入数据：
Flink SQL> create table t1(
> uuid varchar(20),
> name varchar(10),
> age int,
> ts timestamp(3),
> `partition` varchar(20)
> )
> PARTITIONED BY (`partition`)
> with (
> 'connector' = 'hudi',
> 'path' = 'oss://vvr-daily/hudi/t1',
> 'table.type' = 'MERGE_ON_READ',
> 'read.streaming.enabled' = 'true',
> 'read.streaming.check-interval' = '4'
> );
[INFO] Table has been created.
Flink SQL> insert into t1 values
> ('id1','Danny',23,TIMESTAMP '1970-01-01 00:00:01','par1'),
> ('id2','Stephen',33,TIMESTAMP '1970-01-01 00:00:02','par1'),
> ('id3','Julian',53,TIMESTAMP '1970-01-01 00:00:03','par2'),
> ('id4','Fabian',31,TIMESTAMP '1970-01-01 00:00:04','par2'),
> ('id5','Sophia',18,TIMESTAMP '1970-01-01 00:00:05','par3'),
> ('id6','Emma',20,TIMESTAMP '1970-01-01 00:00:06','par3'),
> ('id7','Bob',44,TIMESTAMP '1970-01-01 00:00:07','par4'),
> ('id8','Han',56,TIMESTAMP '1970-01-01 00:00:08','par4');
[INFO] Submitting SQL update statement to the cluster...
[INFO] Table update statement has been successfully submitted to the cluster:
Job ID: 9e1dcd37fd0f8ca77534c30c7d87be2c
这里将 table option read.streaming.enabled 设置为 true，表明通过 streaming 的方式读取表数据；opiton read.streaming.check-interval 指定了 source 监控新的 commits 的间隔为 4s；option table.type 设置表类型为 MERGE_ON_READ，目前只有 MERGE_ON_READ 表支持 streaming 读。
以上操作发生在一个 terminal 中，我们称之为 terminal_1。
从新的 terminal（我们称之为 terminal_2）再次启动 Sql Client，重新创建 t1 表并查询：
Flink SQL> set execution.result-mode=tableau;
[INFO] Session property has been set.
Flink SQL> create table t1(
> uuid varchar(20),
> name varchar(10),
> age int,
> ts timestamp(3),
> `partition` varchar(20)
> )
> PARTITIONED BY (`partition`)
> with (
> 'connector' = 'hudi',
> 'path' = 'oss://vvr-daily/hudi/t1',
> 'table.type' = 'MERGE_ON_READ',
> 'read.streaming.enabled' = 'true',
> 'read.streaming.check-interval' = '4'
> );
[INFO] Table has been created.
Flink SQL> select * from t1;
2021-03-22 18:36:37,042 INFO org.apache.hadoop.conf.Configuration.deprecation [] - mapred.job.map.memory.mb is deprecated. Instead, use mapreduce.map.memory.mb
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| +/- | uuid | name | age | ts | partition |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| + | id2 | Stephen | 33 | 1970-01-01T00:00:02 | par1 |
| + | id1 | Danny | 23 | 1970-01-01T00:00:01 | par1 |
| + | id6 | Emma | 20 | 1970-01-01T00:00:06 | par3 |
| + | id5 | Sophia | 18 | 1970-01-01T00:00:05 | par3 |
| + | id8 | Han | 56 | 1970-01-01T00:00:08 | par4 |
| + | id7 | Bob | 44 | 1970-01-01T00:00:07 | par4 |
| + | id4 | Fabian | 31 | 1970-01-01T00:00:04 | par2 |
| + | id3 | Julian | 53 | 1970-01-01T00:00:03 | par2 |
回到 terminal_1，继续执行 batch mode 的 INSERT 操作：
Flink SQL> insert into t1 values
> ('id1','Danny',27,TIMESTAMP '1970-01-01 00:00:01','par1');
[INFO] Submitting SQL update statement to the cluster...
[INFO] Table update statement has been successfully submitted to the cluster:
Job ID: 2dad24e067b38bc48c3a8f84e793e08b
几秒之后，观察 terminal_2 的输出多了一行：
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| +/- | uuid | name | age | ts | partition |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| + | id2 | Stephen | 33 | 1970-01-01T00:00:02 | par1 |
| + | id1 | Danny | 23 | 1970-01-01T00:00:01 | par1 |
| + | id6 | Emma | 20 | 1970-01-01T00:00:06 | par3 |
| + | id5 | Sophia | 18 | 1970-01-01T00:00:05 | par3 |
| + | id8 | Han | 56 | 1970-01-01T00:00:08 | par4 |
| + | id7 | Bob | 44 | 1970-01-01T00:00:07 | par4 |
| + | id4 | Fabian | 31 | 1970-01-01T00:00:04 | par2 |
| + | id3 | Julian | 53 | 1970-01-01T00:00:03 | par2 |
| + | id1 | Danny | 27 | 1970-01-01T00:00:01 | par1 |
再次在 terminal_1 中执行 INSERT 操作：
Flink SQL> insert into t1 values
> ('id4','Fabian',32,TIMESTAMP '1970-01-01 00:00:04','par2'),
> ('id5','Sophia',19,TIMESTAMP '1970-01-01 00:00:05','par3');
[INFO] Submitting SQL update statement to the cluster...
[INFO] Table update statement has been successfully submitted to the cluster:
Job ID: ecafffda3d294a13b0a945feb9acc8a5
观察 terminal_2 的输出变化：
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| +/- | uuid | name | age | ts | partition |
+-----+----------------------+----------------------+-------------+-------------------------+----------------------+
| + | id2 | Stephen | 33 | 1970-01-01T00:00:02 | par1 |
| + | id1 | Danny | 23 | 1970-01-01T00:00:01 | par1 |
| + | id6 | Emma | 20 | 1970-01-01T00:00:06 | par3 |
| + | id5 | Sophia | 18 | 1970-01-01T00:00:05 | par3 |
| + | id8 | Han | 56 | 1970-01-01T00:00:08 | par4 |
| + | id7 | Bob | 44 | 1970-01-01T00:00:07 | par4 |
| + | id4 | Fabian | 31 | 1970-01-01T00:00:04 | par2 |
| + | id3 | Julian | 53 | 1970-01-01T00:00:03 | par2 |
| + | id1 | Danny | 27 | 1970-01-01T00:00:01 | par1 |
| + | id5 | Sophia | 19 | 1970-01-01T00:00:05 | par3 |
| + | id4 | Fabian | 32 | 1970-01-01T00:00:04 | par2 |
五、总结
通过一些简单的演示，我们发现 HUDI Flink 的集成已经相对完善，读写路径均已覆盖，关于详细的配置，可以参考 Flink SQL Config Options[10]。
Hudi 社区正在积极的推动和 Flink 的深度集成，包括但不限于：
- Flink streaming reader 支持 watermark，实现数据湖/仓的中间计算层 pipeline
- Flink 基于 Hudi 的物化视图，实现分钟级的增量视图，服务于线上的近实时查询
注释： [1] www.alibabacloud.com/help/zh/pro… [2]aws.amazon.com/cn/emr/feat… [3]www.uber.com/ [4]www.slideshare.net/vinothchand… [5]eng.uber.com/uber-big-da… [6]issues.apache.org/jira/browse… [7]ci.apache.org/projects/fl… [8]flink.apache.org/downloads.h… [9]repo.maven.apache.org/maven2/org/… [10]hudi.apache.org/docs/config…
