前言
1. 什么是 Apache Paimon
Apache Paimon (incubating) 是一项**流式数据湖存储**技术,可以为用户提供**高吞吐、低延迟**的数据摄入、流式订阅以及实时查询能力。
Paimon 采用开放的数据格式和技术理念,可以与 Apache Flink / Spark / Trino 等诸多业界主流计算引擎进行对接,共同推进 Streaming Lakehouse 架构的普及和发展。
Paimon 以湖存储的方式基于分布式文件系统管理元数据,并采用开放的** ORC、Parquet、Avro** 文件格式,支持各大主流计算引擎,包括 Flink、Spark、Hive、Trino、Presto。未来会对接更多引擎,包括 Doris 和 Starrocks。
Github:https://github.com/apache/incubator-paimon
以下为快速入门上手Paimon的example:
一、本地环境快速上手
基于paimon 0.4-SNAPSHOT (Flink 1.14.4),Flink版本太低是不支持的,paimon基于最低版本1.14.6,经尝试在Flink1.14.0是不可以的!
paimon-flink-1.14-0.4-20230504.002229-50.jar
1、本地Flink伪集群
需要先下载jar包,并添加至flink的lib中;
根据官网demo,启动flinksql-client,创建catalog,创建表,创建数据源(视图),insert数据到表中。
- 通过 localhost:8081 查看 Flink UI
- 查看filesystem数据、元数据文件
2、IDEA中跑Paimon Demo
pom依赖:
<dependency>
<groupId>org.apache.paimon</groupId>
<artifactId>paimon-flink-1.14</artifactId>
<version>0.4-SNAPSHOT</version>
</dependency>
拉取不到的可以手动添加到本地maven仓库:
mvn install:install-file -DgroupId=org.apache.paimon -DartifactId=paimon-flink-1.14 -Dversion=0.4-SNAPSHOT -Dpackaging=jar -Dfile=D:\software\paimon-flink-1.14-0.4-20230504.002229-50.jar
2.1 代码
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.table.api.TableEnvironment;
import org.apache.flink.table.api.bridge.java.StreamTableEnvironment;
/**
* @Author: YK.Leo
* @Date: 2023-05-14 15:12
* @Version: 1.0
*/
// Succeed at local !!!
public class OfficeDemoV1 {
public static void main(String[] args) throws Exception {
StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(1);
env.enableCheckpointing(10000l);
env.getCheckpointConfig().setCheckpointStorage("file:/D:/tmp/paimon/");
TableEnvironment tableEnv = StreamTableEnvironment.create(env);
// 0. Create a Catalog and a Table
tableEnv.executeSql("CREATE CATALOG my_catalog_api WITH (\n" +
" 'type'='paimon',\n" + // todo: !!!
" 'warehouse'='file:///D:/tmp/paimon'\n" +
")");
tableEnv.executeSql("USE CATALOG my_catalog_api");
tableEnv.executeSql("CREATE TABLE IF NOT EXISTS word_count_api (\n" +
" word STRING PRIMARY KEY NOT ENFORCED,\n" +
" cnt BIGINT\n" +
")");
// 1. Write Data
tableEnv.executeSql("CREATE TEMPORARY TABLE IF NOT EXISTS word_table_api (\n" +
" word STRING\n" +
") WITH (\n" +
" 'connector' = 'datagen',\n" +
" 'fields.word.length' = '1'\n" +
")");
// tableEnv.executeSql("SET 'execution.checkpointing.interval' = '10 s'");
tableEnv.executeSql("INSERT INTO word_count_api SELECT word, COUNT(*) FROM word_table_api GROUP BY word");
env.execute();
}
}
2.2 IDEA中成功运行
3、IDEA中Stream读写
3.1 流写
代码:
package com.study.flink.table.paimon.demo;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.table.api.StatementSet;
import org.apache.flink.table.api.TableEnvironment;
import org.apache.flink.table.api.bridge.java.StreamTableEnvironment;
/**
* @Author: YK.Leo
* @Date: 2023-05-17 11:11
* @Version: 1.0
*/
// succeed at local !!!
public class OfficeStreamsWriteV2 {
public static void main(String[] args) throws Exception {
StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(1);
env.enableCheckpointing(10000L);
env.getCheckpointConfig().setCheckpointStorage("file:/D:/tmp/paimon/");
TableEnvironment tableEnv = StreamTableEnvironment.create(env);
// 0. Create a Catalog and a Table
tableEnv.executeSql("CREATE CATALOG my_catalog_local WITH (\n" +
" 'type'='paimon',\n" + // todo: !!!
" 'warehouse'='file:///D:/tmp/paimon'\n" +
")");
tableEnv.executeSql("USE CATALOG my_catalog_local");
tableEnv.executeSql("CREATE DATABASE IF NOT EXISTS my_catalog_local.local_db");
tableEnv.executeSql("USE local_db");
// drop tbl
tableEnv.executeSql("DROP TABLE IF EXISTS paimon_tbl_streams");
tableEnv.executeSql("CREATE TABLE IF NOT EXISTS paimon_tbl_streams(\n"
+ " uuid bigint,\n"
+ " name VARCHAR(3),\n"
+ " age int,\n"
+ " ts TIMESTAMP(3),\n"
+ " dt VARCHAR(10), \n"
+ " PRIMARY KEY (dt, uuid) NOT ENFORCED \n"
+ ") PARTITIONED BY (dt) \n"
+ " WITH (\n" +
" 'merge-engine' = 'partial-update',\n" +
" 'changelog-producer' = 'full-compaction', \n" +
" 'file.format' = 'orc', \n" +
" 'scan.mode' = 'compacted-full', \n" +
" 'bucket' = '5', \n" +
" 'sink.parallelism' = '5', \n" +
" 'sequence.field' = 'ts' \n" + // todo, to check
")"
);
// datagen ====================================================================
tableEnv.executeSql("CREATE TEMPORARY TABLE IF NOT EXISTS source_A (\n" +
" uuid bigint PRIMARY KEY NOT ENFORCED,\n" +
" `name` VARCHAR(3)," +
" _ts1 TIMESTAMP(3)\n" +
") WITH (\n" +
" 'connector' = 'datagen', \n" +
" 'fields.uuid.kind'='sequence',\n" +
" 'fields.uuid.start'='0', \n" +
" 'fields.uuid.end'='1000000', \n" +
" 'rows-per-second' = '1' \n" +
")");
tableEnv.executeSql("CREATE TEMPORARY TABLE IF NOT EXISTS source_B (\n" +
" uuid bigint PRIMARY KEY NOT ENFORCED,\n" +
" `age` int," +
" _ts2 TIMESTAMP(3)\n" +
") WITH (\n" +
" 'connector' = 'datagen', \n" +
" 'fields.uuid.kind'='sequence',\n" +
" 'fields.uuid.start'='0', \n" +
" 'fields.uuid.end'='1000000', \n" +
" 'rows-per-second' = '1' \n" +
")");
//
//tableEnv.executeSql("insert into paimon_tbl_streams(uuid, name, _ts1) select uuid, concat(name,'_A') as name, _ts1 from source_A");
//tableEnv.executeSql("insert into paimon_tbl_streams(uuid, age, _ts1) select uuid, concat(age,'_B') as age, _ts1 from source_B");
StatementSet statementSet = tableEnv.createStatementSet();
statementSet
.addInsertSql("insert into paimon_tbl_streams(uuid, name, ts, dt) select uuid, name, _ts1 as ts, date_format(_ts1,'yyyy-MM-dd') as dt from source_A")
.addInsertSql("insert into paimon_tbl_streams(uuid, age, dt) select uuid, age, date_format(_ts2,'yyyy-MM-dd') as dt from source_B")
;
statementSet.execute();
// env.execute();
}
}
结果:
如果只有一个流,上述代码完全没有问题【仅作为write demo一个流即可】,两个流会出现“写冲突”问题!
如下:
使用了官网的方法:Dedicated Compaction Job,似乎并没有奏效,至于解决方法请看下文 “**二、进阶:本地(IDEA)多流拼接测试**”;
3.2 流读(toChangeLogStream)
代码:
package com.study.flink.table.paimon.demo;
import org.apache.flink.api.common.functions.FilterFunction;
import org.apache.flink.streaming.api.datastream.DataStream;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.table.api.Schema;
import org.apache.flink.table.api.Table;
import org.apache.flink.table.api.TableEnvironment;
import org.apache.flink.table.api.bridge.java.StreamTableEnvironment;
import org.apache.flink.table.connector.ChangelogMode;
import org.apache.flink.types.Row;
import org.apache.flink.types.RowKind;
import org.apache.logging.log4j.LogManager;
import org.apache.logging.log4j.Logger;
/**
* @Author: YK.Leo
* @Date: 2023-05-15 18:50
* @Version: 1.0
*/
// 流读单表OK!
public class OfficeStreamReadV1 {
public static final Logger LOGGER = LogManager.getLogger(OfficeStreamReadV1.class);
public static void main(String[] args) throws Exception {
StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(1);
env.enableCheckpointing(10000L);
env.getCheckpointConfig().setCheckpointStorage("file:/D:/tmp/paimon/");
TableEnvironment tableEnv = StreamTableEnvironment.create(env);
// 0. Create a Catalog and a Table
tableEnv.executeSql("CREATE CATALOG my_catalog_local WITH (\n" +
" 'type'='paimon',\n" + // todo: !!!
" 'warehouse'='file:///D:/tmp/paimon'\n" +
")");
tableEnv.executeSql("USE CATALOG my_catalog_local");
tableEnv.executeSql("CREATE DATABASE IF NOT EXISTS my_catalog_local.local_db");
tableEnv.executeSql("USE local_db");
// 不需要再次创建表
// convert to DataStream
// Table table = tableEnv.sqlQuery("SELECT * FROM paimon_tbl_streams");
Table table = tableEnv.sqlQuery("SELECT * FROM paimon_tbl_streams WHERE name is not null and age is not null");
// DataStream<Row> dataStream = ((StreamTableEnvironment) tableEnv).toChangelogStream(table);
// todo : doesn't support consuming update and delete changes which is produced by node TableSourceScan
// DataStream<Row> dataStream = ((StreamTableEnvironment) tableEnv).toDataStream(table);
// 剔除 -U 数据(即:更新前的数据不需要重新发送,剔除)!!!
DataStream<Row> dataStream = ((StreamTableEnvironment) tableEnv)
.toChangelogStream(table, Schema.newBuilder().primaryKey("dt","uuid").build(), ChangelogMode.upsert())
.filter(new FilterFunction<Row>() {
@Override
public boolean filter(Row row) throws Exception {
boolean isNoteUpdateBefore = !(row.getKind().equals(RowKind.UPDATE_BEFORE));
if (!isNoteUpdateBefore) {
LOGGER.info("UPDATE_BEFORE: " + row.toString());
}
return isNoteUpdateBefore;
}
})
;
// use this datastream
dataStream.executeAndCollect().forEachRemaining(System.out::println);
env.execute();
}
}
结果:
二、进阶:本地(IDEA)多流拼接测试
要解决的问题:
多个流拥有相同的主键,每个流更新除主键外的部分字段,通过主键完成多流拼接。
note:
如果是两个Flink Job 或者 两个 pipeline 写同一个paimon表,则直接会产生conflict,其中一条流不断exception、重启;
可以使用 “UNION ALL” 将多个流合并为一个流,最终一个Flink job写paimon表;
使用主键表,'merge-engine' = 'partial-update' ;
1、'changelog-producer' = 'full-compaction'
(1)multiWrite代码
package com.study.flink.table.paimon.multi;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.table.api.StatementSet;
import org.apache.flink.table.api.TableEnvironment;
import org.apache.flink.table.api.bridge.java.StreamTableEnvironment;
/**
* @Author: YK.Leo
* @Date: 2023-05-18 10:17
* @Version: 1.0
*/
// Succeed as local !!!
// 而且不会产生conflict,跑5分钟没有任何异常(公司跑几天无异常)! 数据也可以在另一个job流读!
public class MultiStreamsUnionWriteV1 {
public static void main(String[] args) throws Exception {
StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(1);
env.enableCheckpointing(10*1000L);
env.getCheckpointConfig().setCheckpointStorage("file:/D:/tmp/paimon/");
TableEnvironment tableEnv = StreamTableEnvironment.create(env);
// 0. Create a Catalog and a Table
tableEnv.executeSql("CREATE CATALOG my_catalog_local WITH (\n" +
" 'type'='paimon',\n" + // todo: !!!
" 'warehouse'='file:///D:/tmp/paimon'\n" +
")");
tableEnv.executeSql("USE CATALOG my_catalog_local");
tableEnv.executeSql("CREATE DATABASE IF NOT EXISTS my_catalog_local.local_db");
tableEnv.executeSql("USE local_db");
// drop & create tbl
tableEnv.executeSql("DROP TABLE IF EXISTS paimon_tbl_streams");
tableEnv.executeSql("CREATE TABLE IF NOT EXISTS paimon_tbl_streams(\n"
+ " uuid bigint,\n"
+ " name VARCHAR(3),\n"
+ " age int,\n"
+ " ts TIMESTAMP(3),\n"
+ " dt VARCHAR(10), \n"
+ " PRIMARY KEY (dt, uuid) NOT ENFORCED \n"
+ ") PARTITIONED BY (dt) \n"
+ " WITH (\n" +
" 'merge-engine' = 'partial-update',\n" +
" 'changelog-producer' = 'full-compaction', \n" +
" 'file.format' = 'orc', \n" +
" 'scan.mode' = 'compacted-full', \n" +
" 'bucket' = '5', \n" +
" 'sink.parallelism' = '5', \n" +
// " 'write_only' = 'true', \n" +
" 'sequence.field' = 'ts' \n" + // todo, to check
")"
);
// datagen ====================================================================
tableEnv.executeSql("CREATE TEMPORARY TABLE IF NOT EXISTS source_A (\n" +
" uuid bigint PRIMARY KEY NOT ENFORCED,\n" +
" `name` VARCHAR(3)," +
" _ts1 TIMESTAMP(3)\n" +
") WITH (\n" +
" 'connector' = 'datagen', \n" +
" 'fields.uuid.kind'='sequence',\n" +
" 'fields.uuid.start'='0', \n" +
" 'fields.uuid.end'='1000000', \n" +
" 'rows-per-second' = '1' \n" +
")");
tableEnv.executeSql("CREATE TEMPORARY TABLE IF NOT EXISTS source_B (\n" +
" uuid bigint PRIMARY KEY NOT ENFORCED,\n" +
" `age` int," +
" _ts2 TIMESTAMP(3)\n" +
") WITH (\n" +
" 'connector' = 'datagen', \n" +
" 'fields.uuid.kind'='sequence',\n" +
" 'fields.uuid.start'='0', \n" +
" 'fields.uuid.end'='1000000', \n" +
" 'rows-per-second' = '1' \n" +
")");
//
StatementSet statementSet = tableEnv.createStatementSet();
String sqlText = "INSERT INTO paimon_tbl_streams(uuid, name, age, ts, dt) \n" +
"select uuid, name, cast(null as int) as age, _ts1 as ts, date_format(_ts1,'yyyy-MM-dd') as dt from source_A \n" +
"UNION ALL \n" +
"select uuid, cast(null as string) as name, age, _ts2 as ts, date_format(_ts2,'yyyy-MM-dd') as dt from source_B"
;
statementSet.addInsertSql(sqlText);
statementSet.execute();
}
}
读代码同上。
(2)读延迟
即:从client数据落到paimon,完成与server的join,再到被Flink-paimon流读到的时间延迟;
**分钟级别延迟**!
2、'changelog-producer' = 'lookup'
读写同上,建表时修改参数即可: changelog-producer='lookup',与此匹配的scan-mode需要分别配置为 'latest' ;
lookup延迟性可能会更低,但是数据质量有待验证。
note:
经测试,在企业生产环境中full-compaction模式目前一切稳定(两条join的流QPS约3K左右,延迟2-3分钟)。
99.9%的数据延迟在2-3分钟;
(multiWrite的checkpoint间隔为60s时)
三、可能遇到的问题
- Caused by: java.lang.ClassCastException: org.codehaus.janino.CompilerFactory cannot be cast to org.codehaus.commons.compiler.ICompilerFactory
原因:org.codehaus.janino 依赖冲突,
办法:全部exclude掉
<exclude>org.codehaus.janino:*</exclude>
- Caused by: java.lang.ClassNotFoundException: org.apache.flink.util.function.SerializableFunction
原因:Flink steaming版本与Flink table版本不一致 或 确实相关依赖 (这里是paimon依赖的flink版本最低为1.14.6,与1.14.0的flink不兼容)
办法:升级Flink版本到1.14.4以上
参考Flink配置:Configuration | Apache Flink
- Caused by: java.util.ServiceConfigurationError: org.apache.flink.table.factories.Factory: Provider org.apache.flink.table.store.connector.TableStoreManagedFactory not found
在项目的META-INF/services路径下添加 Factory 文件(这样才能匹配Flink的CatalogFactory,才能创建catalog)
- Caused by: org.apache.flink.client.program.ProgramInvocationException: The main method caused an error: No operators defined in streaming topology. Cannot execute.
已经存在tableEnv.executeSql 或者 statementSet.execute() 时就不需要再 env.execute() 了!
Flink SQL不能直接使用null as,需要写成 cast(null as data_type), 如 cast(null as string);
如果创建paimon分区表,必须要把分区字段放在主键中!,否则建表报错:
四、展望
如果有数据格式:
**主键 stream_client stream_server ts **
1001 null a 1
1001 A null 2
1001 B null 3
按照paimon官方的实现,使用主键表的partial update进行多流拼接会被拼接为如下结果:
1001 B a 3;
即:**主键会被去重**(取每个流里边最新的一条),如果想要保留 stream_client 的全部数据,官方源码实现不了,需要进行改造!
我们已经改造并实现了非去重的效果,后续出一篇专门的文章阐述一下改造思路和方法。
想象:
stream_client为客户端数据,请求一次服务之后,可以上下滑动屏幕(或者进入后回退),使某个商品产生多次曝光(但不会多次请求server端);此时 client 端产生了多条数据,server端只有一条数据。但是,client端多次的曝光/点击是可以反应用户对某个商品的感兴趣程度的,是有意义的数据,不应该被去重掉!
【未完待续...】
版权归原作者 Leonardo_KY 所有, 如有侵权,请联系我们删除。