Flink Upsert Kafka SQL Connector 介绍

一 前言

在某些场景中，比方GROUP BY聚合之后的后果，须要去更新之前的结果值。这个时候，须要将 Kafka 记录的 key 当成主键解决，用来确定一条数据是应该作为插入、删除还是更新记录来解决。在 Flink1.11 中，能够通过 flink-cdc-connectors 项目提供的 changelog-json format 来实现该性能。

在 Flink1.12 版本中， 新增了一个 upsert connector(upsert-kafka)，该 connector 扩大自现有的 Kafka connector，工作在 upsert 模式（FLIP-149）下。新的 upsert-kafka connector 既能够作为 source 应用，也能够作为 sink 应用，并且提供了与现有的 kafka connector 雷同的基本功能和持久性保障，因为两者之间复用了大部分代码。

二 upsert kafka connector

Upsert Kafka Connector容许用户以upsert的形式从Kafka主题读取数据或将数据写入Kafka主题。

作为 source，upsert-kafka 连接器生产 changelog 流，其中每条数据记录代表一个更新或删除事件。更准确地说，数据记录中的 value 被解释为同一 key 的最后一个 value 的 UPDATE，如果有这个 key（如果不存在相应的 key，则该更新被视为 INSERT）。用表来类比，changelog 流中的数据记录被解释为 UPSERT，也称为 INSERT/UPDATE，因为任何具有相同 key 的现有行都被覆盖。另外，value 为空的消息将会被视作为 DELETE 消息。

作为 sink，upsert-kafka 连接器可以消费 changelog 流。它会将 INSERT/UPDATE_AFTER 数据作为正常的 Kafka 消息写入，并将 DELETE 数据以 value 为空的 Kafka 消息写入（表示对应 key 的消息被删除）。Flink 将根据主键列的值对数据进行分区，从而保证主键上的消息有序，因此同一主键上的更新/删除消息将落在同一分区中。

其中每条数据记录代表一个更新或删除事件，原理如下：

• Kafka Topic中存在相应的Key，则以UPDATE操作将Key的值更新为数据记录中的Value。
• Kafka Topic中不存在相应的Key，则以INSERT操作将Key的值写入Kafka Topic。
• Key对应的Value为空，会被视作DELETE操作。

三 案例

3.1 kafka 处理后写入kafka

3.1.1 创建kafka topic

$ kafka-topics --create --topic user-behavior --partitions 3 --replication-factor 2 --bootstrap-server cdh68:9092,cdh69:9092,cdh70:9092
$ kafka-topics --create --topic after-user-behavior --partitions 3 --replication-factor 2 --bootstrap-server cdh68:9092,cdh69:9092,cdh70:9092
$ kafka-console-producer --topic user-behavior --broker-list cdh68:9092,cdh69:9092,cdh70:9092
$ kafka-console-consumer --topic user-behavior --from-beginning --group test-user --bootstrap-server cdh68:9092,cdh69:9092,cdh70:9092      
$ kafka-console-consumer --topic after-user-behavior --from-beginning --group test --bootstrap-server cdh68:9092,cdh69:9092,cdh70:9092

3.2 Flink SQL

3.2.1 source

%flink.ssql
drop table if exists user_behavior;
CREATE TABLE user_behavior (
    id BIGINT,
    name STRING,
    flag STRING
) WITH (
    'connector' = 'kafka',  -- 使用 kafka connector
    'topic' = 'user-behavior',  -- kafka topic
    'properties.group.id'='cdc', -- 消费者组
    'scan.startup.mode' = 'latest-offset',  -- 从起始 offset 开始读取
    'json.fail-on-missing-field' = 'false',
    'json.ignore-parse-errors' = 'true',
    'properties.bootstrap.servers' = 'cdh68:9092,cdh69:9092,cdh70:9092',  -- kafka broker 地址
    'format' = 'json'  -- 数据源格式为 json
);

3.2.2 sink

%flink.ssql
drop table if exists after_user_behavior;
CREATE TABLE after_user_behavior (
  name STRING,
  pv BIGINT,
  PRIMARY KEY (name) NOT ENFORCED
) WITH (
  'connector' = 'upsert-kafka',
  'topic' = 'after-user-behavior',
  'properties.bootstrap.servers' = 'cdh68:9092,cdh69:9092,cdh70:9092',
  'value.json.fail-on-missing-field' = 'false',
  'key.json.ignore-parse-errors' = 'true',
  'key.format' = 'json',
  'value.format' = 'json'
);

一定要设置主键 Primar要使用 upsert-kafka connector，DDL语句中，一定要设置 PRIMARY KEY 主键，并为键（key.format）和值（value.format）指定序列化反序列化格式。

当数据源端进行了增删改，对应的 pv 结果就会同步更新，这就是 upsert kafka 的魅力。

这是基于kafka的统计计算，前提条件是 topic user-behavior中的数据是 changelog 格式的。

3.2.3 transform

%flink.ssql
INSERT INTO after_user_behavior
SELECT
  name,
  COUNT(*)
FROM user_behavior 
GROUP BY name;

注意：after_user_behavior 必须为 upsert-kafka connector

如果after_user_behavior为 kafka connector,执行此语句则会报如下错误：

java.io.IOException: org.apache.flink.table.api.TableException: Table sink 'default_catalog.default_database.after_user_behavior' doesn't support consuming update changes which is produced by node GroupAggregate(groupBy=[name], select=[name, COUNT(*) AS EXPR$1])

因为语句SELECT name, COUNT(*) FROM user_behavior GROUP BY name; 通过group by后数据是不断更新变化的，因此只能用 upsert-kafka connector。

3.3 输出结果

3.3.1 kafka user-behavior producer

[song@cdh68 ~]$ kafka-console-producer --topic user-behavior --broker-list cdh68:9092,cdh69:9092,cdh70:9092
>{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"INSERT","table":"user","ts":6852139698555588608}
>{"schema":"presto","flag":false,"name":"Lucy","id":"67","type":"INSERT","table":"user","ts":6852139698555588608}
>{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"INSERT","table":"info","ts":6852139698555588608}
>{"schema":"presto","flag":false,"name":"Lucy","id":"67","type":"INSERT","table":"info","ts":6852139698555588608}
>{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"INSERT","table":"user","ts":6852139698555588608}
>{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"INSERT","table":"user","ts":6852139698555588608}
>{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"UPDATE","table":"user","ts":6852139698555588608}
>{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"DELETE","table":"user","ts":6852139698555588608}

topic user-behavior中的数据是 changelog 格式的。

3.3.2 kafka user-behavior consumer

[song@cdh70 ~]$ kafka-console-consumer --topic user-behavior --group test-user --bootstrap-server cdh68:9092,cdh69:9092,cdh70:9092
{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"INSERT","table":"user","ts":6852139698555588608}
{"schema":"presto","flag":false,"name":"Lucy","id":"67","type":"INSERT","table":"user","ts":6852139698555588608}
{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"INSERT","table":"info","ts":6852139698555588608}
{"schema":"presto","flag":false,"name":"Lucy","id":"67","type":"INSERT","table":"info","ts":6852139698555588608}
{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"INSERT","table":"user","ts":6852139698555588608}
{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"INSERT","table":"user","ts":6852139698555588608}
{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"UPDATE","table":"user","ts":6852139698555588608}
{"schema":"presto","flag":false,"name":"Mars","id":"85","type":"DELETE","table":"user","ts":6852139698555588608}

3.3.3 kafka after-user-behavior consumer

[song@cdh69 ~]$ kafka-console-consumer --topic after-user-behavior --group test --bootstrap-server cdh68:9092,cdh69:9092,cdh70:9092
{"name":"Mars","pv":1}
{"name":"Lucy","pv":1}
{"name":"Mars","pv":2}
{"name":"Lucy","pv":2}
{"name":"Mars","pv":3}
{"name":"Mars","pv":4}
{"name":"Mars","pv":5}
{"name":"Mars","pv":6}

3.3.4 FlinkSQL user_behavior

从此结果可以看出 kafka 和 upsert-kafka 的区别：

kafka 的结果则显示所有数据，upsert-kafka则显示更新后的最新数据。

3.3.5 FlinkSQL alfter_user_behavior

此结果是动态变化的，变化与kafka after-user-behavior consumer相同。

可见，upsert-kafka 表存储了所有变化的数据，但是读取时，只读取最新的数据。

3.2 flink-pageviews-demo

https://github.com/fsk119/flink-pageviews-demo

3.2.1 测试数据准备

在 Mysql 中执行以下命令：

CREATE DATABASE flink;
USE flink;

CREATE TABLE users (
  user_id BIGINT,
  user_name VARCHAR(1000),
  region VARCHAR(1000)
);

INSERT INTO users VALUES 
(1, 'Timo', 'Berlin'),
(2, 'Tom', 'Beijing'),
(3, 'Apple', 'Beijing');

现在，我们利用Sql client在Flink中创建相应的表。

CREATE TABLE users (
  user_id BIGINT,
  user_name STRING,
  region STRING
) WITH (
  'connector' = 'mysql-cdc',
  'hostname' = 'localhost',
  'database-name' = 'flink',
  'table-name' = 'users',
  'username' = 'root',
  'password' = '123456'
);

CREATE TABLE pageviews (
  user_id BIGINT,
  page_id BIGINT,
  view_time TIMESTAMP(3),
  proctime AS PROCTIME()
) WITH (
  'connector' = 'kafka',
  'topic' = 'pageviews',
  'properties.bootstrap.servers' = 'localhost:9092',
  'scan.startup.mode' = 'earliest-offset',
  'format' = 'json'
);

并利用Flink 往 Kafka中灌入相应的数据

INSERT INTO pageviews VALUES
  (1, 101, TO_TIMESTAMP('2020-11-23 15:00:00')),
  (2, 104, TO_TIMESTAMP('2020-11-23 15:00:01.00'));

3.2.2 将 left join 结果写入 kafka

我们首先测试是否能将Left join的结果灌入到 Kafka 之中。

首先，我们在 Sql client 中创建相应的表

CREATE TABLE enriched_pageviews (
  user_id BIGINT,
  user_region STRING,
  page_id BIGINT,
  view_time TIMESTAMP(3),
  WATERMARK FOR view_time as view_time - INTERVAL '5' SECOND,
  PRIMARY KEY (user_id, page_id) NOT ENFORCED
) WITH (
  'connector' = 'upsert-kafka',
  'topic' = 'enriched_pageviews',
  'properties.bootstrap.servers' = 'localhost:9092',
  'key.format' = 'json',
  'value.format' = 'json'
);

并利用以下语句将left join的结果插入到kafka对应的topic之中。

INSERT INTO enriched_pageviews
SELECT pageviews.user_id, region, pageviews.page_id, pageviews.view_time
FROM pageviews
LEFT JOIN users ON pageviews.user_id = users.user_id;

利用以下命令，我们可以打印topic内的数据kafka-console-consumer.sh --bootstrap-server kafka:9094 --topic "enriched_pageviews" --from-beginning --property print.key=true

#预期结果
{"user_id":1,"page_id":101}	{"user_id":1,"user_region":null,"page_id":101,"view_time":"2020-11-23 15:00:00"}
{"user_id":2,"page_id":104}	{"user_id":2,"user_region":null,"page_id":104,"view_time":"2020-11-23 15:00:01"}
{"user_id":1,"page_id":101}	null
{"user_id":1,"page_id":101}	{"user_id":1,"user_region":"Berlin","page_id":101,"view_time":"2020-11-23 15:00:00"}
{"user_id":2,"page_id":104}	null
{"user_id":2,"page_id":104}	{"user_id":2,"user_region":"Beijing","page_id":104,"view_time":"2020-11-23 15:00:01"}

Left join中，右流发现左流没有join上但已经发射了，此时会发送DELETE消息，而非UPDATE-BEFORE消息清理之前发送的消息。详见org.apache.flink.table.runtime.operators.join.stream.StreamingJoinOperator#processElement

我们可以进一步在mysql中删除或者修改一些数据，来观察进一步的变化。

UPDATE users SET region = 'Beijing' WHERE user_id = 1;

DELETE FROM users WHERE user_id = 1;

3.2.3 将聚合结果写入kafka

我们进一步测试将聚合的结果写入到 Kafka 之中。

在Sql client 中构建以下表

CREATE TABLE pageviews_per_region (
  user_region STRING,
  cnt BIGINT,
  PRIMARY KEY (user_region) NOT ENFORCED
) WITH (
  'connector' = 'upsert-kafka',
  'topic' = 'pageviews_per_region',
  'properties.bootstrap.servers' = 'localhost:9092',
  'key.format' = 'json',
  'value.format' = 'json'
)

我们再用以下命令将数据插入到upsert-kafka之中。

INSERT INTO pageviews_per_region
SELECT
  user_region,
  COUNT(*)
FROM enriched_pageviews
WHERE user_region is not null
GROUP BY user_region;

我们可以通过以下命令查看 Kafka 中对应的数据

./kafka-console-consumer.sh --bootstrap-server kafka:9094 --topic "pageviews_per_region" --from-beginning --property print.key=true

# 预期结果
{"user_region":"Berlin"}	{"user_region":"Berlin","cnt":1}
{"user_region":"Beijing"}	{"user_region":"Beijing","cnt":1}
{"user_region":"Berlin"}	null
{"user_region":"Beijing"}	{"user_region":"Beijing","cnt":2}
{"user_region":"Beijing"}	{"user_region":"Beijing","cnt":1}

---