實(shí)時(shí)數(shù)據(jù)湖：Flink CDC流式寫入Hudi
點(diǎn)擊上方藍(lán)色字體，選擇“設(shè)為星標(biāo)”
回復(fù)"面試"獲取更多驚喜
1. 環(huán)境準(zhǔn)備
?Flink 1.12.2_2.11?Hudi 0.9.0-SNAPSHOT(master分支)?Spark 2.4.5、Hadoop 3.1.3、Hive 3.1.2
2. Flink CDC寫入Hudi
MySQL建表語句如下
create table users(id bigint auto_increment primary key,name varchar(20) null,birthday timestamp default CURRENT_TIMESTAMP not null,ts timestamp default CURRENT_TIMESTAMP not null);// 隨意插入幾條數(shù)據(jù)insert into users (name) values ('hello');insert into users (name) values ('world');insert into users (name) values ('iceberg');insert into users (id,name) values (4,'spark');insert into users (name) values ('hudi');select * from users;update users set name = 'hello spark' where id = 5;delete from users where id = 5;
啟動(dòng)sql-client
$FLINK_HOME/bin/sql-client.sh embedded//1.創(chuàng)建 mysql-cdcCREATE TABLE mysql_users (id BIGINT PRIMARY KEY NOT ENFORCED ,name STRING,birthday TIMESTAMP(3),ts TIMESTAMP(3)) WITH ('connector' = 'mysql-cdc','hostname' = 'localhost','port' = '3306','username' = 'root','password' = '123456','server-time-zone' = 'Asia/Shanghai','database-name' = 'mydb','table-name' = 'users');// 2.創(chuàng)建hudi表CREATE TABLE hudi_users2(id BIGINT PRIMARY KEY NOT ENFORCED,name STRING,birthday TIMESTAMP(3),ts TIMESTAMP(3),`partition` VARCHAR(20)) PARTITIONED BY (`partition`) WITH ('connector' = 'hudi','table.type' = 'MERGE_ON_READ','path' = 'hdfs://localhost:9000/hudi/hudi_users2','read.streaming.enabled' = 'true','read.streaming.check-interval' = '1');//3.mysql-cdc 寫入hudi ，會(huì)提交有一個(gè)flink任務(wù)INSERT INTO hudi_users2 SELECT *, DATE_FORMAT(birthday, 'yyyyMMdd') FROM mysql_users;
Flink任務(wù)提交成功后可以查看任務(wù)界面
同時(shí)可以查看HDFS里的Hudi數(shù)據(jù)路徑，這里需要等Flink 5次checkpoint（默認(rèn)配置可修改）之后才能查看到這些目錄，一開始只有.hoodie一個(gè)文件夾
在MySQL執(zhí)行insert、update、delete等操作，當(dāng)進(jìn)行compaction生成parquet文件后就可以用hive/spark-sql/presto(本文只做了hive和spark-sql的測試)進(jìn)行查詢，這里需要注意下：如果沒有生成parquet文件，我們建的parquet表是查詢不出數(shù)據(jù)的。
3. Hive查詢Hudi表
cd $HIVE_HOMEmkdir auxlib
然后將hudi-hadoop-mr-bundle-0.9.0-SNAPSHOT.jar拷貝過來
使用beeline登錄hive
beeline -u jdbc:hive2://localhost:10000 -n hadoop hadoop創(chuàng)建外部表關(guān)聯(lián)Hudi路徑，有兩種建表方式
方式一：INPUTFORMAT是org.apache.hudi.hadoop.HoodieParquetInputFormat這種方式只會(huì)查詢出來parquet數(shù)據(jù)文件中的內(nèi)容，但是剛剛更新或者刪除的數(shù)據(jù)不能查出來// 創(chuàng)建外部表CREATE EXTERNAL TABLE `hudi_users_2`(`_hoodie_commit_time` string,`_hoodie_commit_seqno` string,`_hoodie_record_key` string,`_hoodie_partition_path` string,`_hoodie_file_name` string,`id` bigint,`name` string,`birthday` bigint,`ts` bigint)PARTITIONED BY (`partition` string)ROW FORMAT SERDE'org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe'STORED AS INPUTFORMAT'org.apache.hudi.hadoop.HoodieParquetInputFormat'OUTPUTFORMAT'org.apache.hadoop.hive.ql.io.parquet.MapredParquetOutputFormat'LOCATION'hdfs://localhost:9000/hudi/hudi_users2';方式二：INPUTFORMAT是org.apache.hudi.hadoop.realtime.HoodieParquetRealtimeInputFormat// 這種方式是能夠?qū)崟r(shí)讀出來寫入的數(shù)據(jù)，也就是Merge On Write，會(huì)將基于Parquet的基礎(chǔ)列式文件、和基于行的Avro日志文件合并在一起呈現(xiàn)給用戶。CREATE EXTERNAL TABLE `hudi_users_2_mor`(`_hoodie_commit_time` string,`_hoodie_commit_seqno` string,`_hoodie_record_key` string,`_hoodie_partition_path` string,`_hoodie_file_name` string,`id` bigint,`name` string,`birthday` bigint,`ts` bigint)PARTITIONED BY (`partition` string)ROW FORMAT SERDE'org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe'STORED AS INPUTFORMAT'org.apache.hudi.hadoop.realtime.HoodieParquetRealtimeInputFormat'OUTPUTFORMAT'org.apache.hadoop.hive.ql.io.parquet.MapredParquetOutputFormat'LOCATION'hdfs://localhost:9000/hudi/hudi_users2';// 添加分區(qū)alter table hudi_users_2 add if not exists partition(`partition`='20210414') location 'hdfs://localhost:9000/hudi/hudi_users2/20210414';alter table hudi_users_2_mor add if not exists partition(`partition`='20210414') location 'hdfs://localhost:9000/hudi/hudi_users2/20210414';// 查詢分區(qū)的數(shù)據(jù)select * from hudi_users_2 where `partition`=20210414;select * from hudi_users_2_mor where `partition`=20210414;
INPUTFORMAT是org.apache.hudi.hadoop.realtime.HoodieParquetRealtimeInputFormat格式的表在hive3.1.2里面是不能夠執(zhí)行統(tǒng)計(jì)操作的
執(zhí)行select count(1) from hudi_users3_mor where partition='20210414';
查看hive日志 tail -fn 100 hiveserver2.log
需要進(jìn)行如下設(shè)置：set hive.input.format = org.apache.hudi.hadoop.hive.HoodieCombineHiveInputFormat ;具體原因參照這個(gè)issue：https://github.com/apache/hudi/issues/2813，或者阿里云技術(shù)文檔：https://help.aliyun.com/document_detail/193310.html?utm_content=g_1000230851&spm=5176.20966629.toubu.3.f2991ddcpxxvD1#title-ves-82n-odd
再執(zhí)行一遍依舊報(bào)錯(cuò)
但是在本地用hive-2.3.8執(zhí)行成功了，社群里面的同學(xué)測試1.1版本的也報(bào)同樣的錯(cuò)誤，目前猜測是hive版本兼容性有關(guān)
4. Spark-SQL查詢Hudi表
將hudi-spark-bundle_2.11-0.9.0-SNAPSHOT.jar拷貝到$SPAKR_HOME/jars，每個(gè)節(jié)點(diǎn)都拷貝一份
將hudi-hadoop-mr-bundle-0.9.0-SNAPSHOT.jar拷貝到$HADOOP_HOME/share/hadoop/hdfs下，每個(gè)節(jié)點(diǎn)都拷貝一份，然后重啟hadoop
創(chuàng)建表，同樣有兩種方式
CREATE EXTERNAL TABLE `hudi_users3_spark`(`_hoodie_commit_time` string,`_hoodie_commit_seqno` string,`_hoodie_record_key` string,`_hoodie_partition_path` string,`_hoodie_file_name` string,`id` bigint,`name` string,`birthday` bigint,`ts` bigint)PARTITIONED BY (`partition` string)ROW FORMAT SERDE'org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe'STORED AS INPUTFORMAT'org.apache.hudi.hadoop.HoodieParquetInputFormat'OUTPUTFORMAT'org.apache.hadoop.hive.ql.io.parquet.MapredParquetOutputFormat'LOCATION'hdfs://localhost:9000/hudi/hudi_users3';alter table hudi_users3_spark add if not exists partition(`partition`='20210414') location 'hdfs://localhost:9000/hudi/hudi_users3/20210414';select * from hudi_users3_spark where `partition`='20210414';// 創(chuàng)建可以實(shí)時(shí)讀表數(shù)據(jù)的格式CREATE EXTERNAL TABLE `hudi_users3_spark_mor`(`_hoodie_commit_time` string,`_hoodie_commit_seqno` string,`_hoodie_record_key` string,`_hoodie_partition_path` string,`_hoodie_file_name` string,`id` bigint,`name` string,`birthday` bigint,`ts` bigint)PARTITIONED BY (`partition` string)ROW FORMAT SERDE'org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe'STORED AS INPUTFORMAT'org.apache.hudi.hadoop.HoodieParquetInputFormat'OUTPUTFORMAT'org.apache.hudi.hadoop.realtime.HoodieParquetRealtimeInputFormat'LOCATION'hdfs://localhost:9000/hudi/hudi_users3';alter table hudi_users3_spark_mor add if not exists partition(`partition`='20210414') location 'hdfs://localhost:9000/hudi/hudi_users3/20210414';select * from hudi_users3_spark_mor where `partition`='20210414';
如果Spark-SQL讀取實(shí)時(shí)Hudi數(shù)據(jù)，必須進(jìn)行如下設(shè)置set spark.sql.hive.convertMetastoreParquet=false;
這里需要注意如果創(chuàng)建表的時(shí)候字段類型不對會(huì)報(bào)錯(cuò)，比如
CREATE EXTERNAL TABLE `hudi_users3_spark_mor`(`_hoodie_commit_time` string,`_hoodie_commit_seqno` string,`_hoodie_record_key` string,`_hoodie_partition_path` string,`_hoodie_file_name` string,`id` string,`name` string,`birthday` string,`ts` string)PARTITIONED BY (`partition` string)ROW FORMAT SERDE'org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe'STORED AS INPUTFORMAT'org.apache.hudi.hadoop.HoodieParquetInputFormat'OUTPUTFORMAT'org.apache.hudi.hadoop.realtime.HoodieParquetRealtimeInputFormat'LOCATION'hdfs://localhost:9000/hudi/hudi_users3';
id 、ts、birthday都設(shè)置為String，會(huì)報(bào)下面錯(cuò)誤。Spark-SQL想讀取Hudi數(shù)據(jù)，字段類型需要嚴(yán)格匹配
5. 后續(xù)
目前使用小規(guī)模數(shù)據(jù)測試Flink CDC寫入Hudi，后面我們準(zhǔn)備用生產(chǎn)數(shù)據(jù)來走一波，看看Flink-CDC寫入Hudi的性能和穩(wěn)定性。
你好，我是王知無，一個(gè)大數(shù)據(jù)領(lǐng)域的硬核原創(chuàng)作者。
做過后端架構(gòu)、數(shù)據(jù)中間件、數(shù)據(jù)平臺(tái)&架構(gòu)、算法工程化。
專注大數(shù)據(jù)領(lǐng)域?qū)崟r(shí)動(dòng)態(tài)&技術(shù)提升&個(gè)人成長&職場進(jìn)階，歡迎關(guān)注。