Creating a FlinkServer Job to Write Data to a Hudi Table

Updated on 2024-12-13 GMT+08:00

View PDF

This section applies to MRS 3.1.2 or later clusters.

Scenario

This section describes how to interconnect FlinkServer with Hudi through Flink SQL jobs. When you use Flink SQL to read data from or write data to Hudi, the TINYINT, SMALLINT, and TIME types cannot be defined.

Table 1 lists the read and write operations supported by Flink on Hudi COW and MOR tables.

**Table 1** Flink SQL read and write operations on Hudi tables
Flink SQL	COW Table	MOR Table
Batch write	Supported	Supported
Batch read	Supported	Supported
Stream write	Supported	Supported
Stream read	Supported	Supported

Prerequisites

The HDFS, Yarn, Flink, and Hudi services have been installed in a cluster.
The client that contains the Hudi service has been installed, for example, in the /opt/client directory.
Flink 1.12.2 or later and Hudi 0.9.0 or later are required.
You have created a user assigned with the FlinkServer Admin Privilege (for example, flink_admin) for accessing the Flink web UI by referring to Creating a FlinkServer Role. The user has been added to the hadoop, hive, and kafkaadmin user groups and granted the Manager_administrator role.

Creating a Job

Log in to Manager as user flink_admin and choose Cluster > Services > Flink. In the Basic Information area, click the link on the right of Flink WebUI to access the Flink web UI.

Create a Flink SQL job by referring to Creating a FlinkServer Job. On the job development page, configure the job as follows: Enter the SQL statement. After the SQL statement passes the verification, start the job. The following SQL examples are added as three jobs and run in sequence.

In Basic Parameter, select Enable CheckPoint, set Time Interval(ms) to 60000, and retain the default value for Mode.

NOTE:

CheckPoint should be enabled on the Flink web UI because data is written to a Hudi table only when a Flink SQL job triggers CheckPoint. Adjust the CheckPoint interval based on service requirements. You are advised to set the interval to a large number.
If the CheckPoint interval is too short, job exceptions may occur due to untimely data updates. It is recommended that the CheckPoint interval be configured at the minute level.
Asynchronous compaction is required when a Flink SQL job writes an MOR table. For details about the parameter for controlling the compaction interval, visit Hudi official website https://hudi.apache.org/docs/configurations.html.
The following content is available in MRS 3.2.1 or later. By default, Hudi table write statements use Flink status index. To use the bucket index, add the following parameters to the statement:
```
'index.type'='BUCKET',
'hoodie.bucket.index.num.buckets'='Number of buckets in each partition of a Hudi table'
'hoodie.bucket.index.hash.field'='recordkey.field'
```
- hoodie.bucket.index.num.buckets: Number of buckets in each partition of a Hudi table. Data in each partition is stored in each bucket in hash mode. This parameter cannot be modified after being set during table creation or data writing for the first time. Otherwise, an exception occurs during data update.
- hoodie.bucket.index.hash.field: Field for calculating the hash value during bucketing. The field must be a subset of the primary key. The default value is the primary key of the Hudi table. If this parameter is left blank, the default value recordkey.field is used.
The following content is available in MRS 3.2.1 or later. The same Hudi table can be written by the bucket indexes across the Flink and Spark engines.

Job 1: This Flink SQL job writes data to an MOR table in streams.

CREATE TABLE stream_mor(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) PARTITIONED BY (`p`) WITH (
'connector' = 'hudi',
'path' = 'hdfs://hacluster/tmp/hudi/stream_mor',
'table.type' = 'MERGE_ON_READ',
'hoodie.datasource.write.recordkey.field' = 'uuid',
'write.precombine.field' = 'ts',
'write.tasks' = '4'
);

CREATE TABLE kafka(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) WITH (
'connector' = 'kafka',
'topic' = 'writehudi',
'properties.bootstrap.servers' = 'IP address of the Kafka broker instance:Kafka port number',
'properties.group.id' = 'testGroup1',
'scan.startup.mode' = 'latest-offset',
'format' = 'json',
'properties.sasl.kerberos.service.name' = 'kafka',--This parameter is not required for clusters in normal mode. Delete the comma (,) in the previous line.
'properties.security.protocol' = 'SASL_PLAINTEXT',--This parameter is not required for clusters in normal mode.
'properties.kerberos.domain.name' = 'hadoop.System domain name'--This parameter is not required for clusters in normal mode.
);

insert into
stream_mor
select
*
from
kafka;

Job 2: This Flink SQL job writes data to a COW table in streams.

CREATE TABLE stream_write_cow(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) PARTITIONED BY (`p`) WITH (
'connector' = 'hudi',
'path' = 'hdfs://hacluster/tmp/hudi/stream_cow',
'hoodie.datasource.write.recordkey.field' = 'uuid',
'write.precombine.field' = 'ts',
'write.tasks' = '4'
);

CREATE TABLE kafka(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) WITH (
'connector' = 'kafka',
'topic' = 'writehudi',
'properties.bootstrap.servers' = 'IP address of the Kafka broker instance:Kafka port number',
'properties.group.id' = 'testGroup1',
'scan.startup.mode' = 'latest-offset',
'format' = 'json',
'properties.sasl.kerberos.service.name' = 'kafka',--This parameter is not required for clusters in normal mode. Delete the comma (,) in the previous line.
'properties.security.protocol' = 'SASL_PLAINTEXT',--This parameter is not required for clusters in normal mode.
'properties.kerberos.domain.name' = 'hadoop.System domain name'--This parameter is not required for clusters in normal mode.
);

insert into
stream_write_cow
select
*
from
kafka;

Job 3: This Flink SQL job reads MOR and COW tables in streams, merges data, and outputs the merged data to Kafka. Verify the SQL statement of job 3 and start it after job 1 and job 2 are started and their status is running. Otherwise, an error message may be displayed during SQL verification, indicating that the Hudi table directory cannot be found.

CREATE TABLE stream_mor(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) PARTITIONED BY (`p`) WITH (
'connector' = 'hudi',  
'path' = 'hdfs://hacluster/tmp/hudi/stream_mor',  
'table.type' = 'MERGE_ON_READ',
'hoodie.datasource.write.recordkey.field' = 'uuid',
'write.precombine.field' = 'ts',
'read.tasks' = '4',
'read.streaming.enabled' = 'true',
'read.streaming.check-interval' = '5',
'read.streaming.start-commit' = 'earliest'
);
CREATE TABLE stream_write_cow(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) PARTITIONED BY (`p`) WITH (
'connector' = 'hudi',
'path' = 'hdfs://hacluster/tmp/hudi/stream_cow',
'hoodie.datasource.write.recordkey.field' = 'uuid',
'write.precombine.field' = 'ts',
'read.tasks' = '4',
'read.streaming.enabled' = 'true',
'read.streaming.check-interval' = '5',
'read.streaming.start-commit' = 'earliest'
);

CREATE TABLE kafka(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) WITH (
'connector' = 'kafka',
'topic' = 'readhudi',
'properties.bootstrap.servers' = 'IP address of the Kafka broker instance:Kafka port',
'properties.group.id' = 'testGroup1',
'scan.startup.mode' = 'latest-offset',
'format' = 'json',
'properties.sasl.kerberos.service.name' = 'kafka',--This parameter is not required for clusters in normal mode. Delete the comma (,) in the previous line.
'properties.security.protocol' = 'SASL_PLAINTEXT',--This parameter is not required for clusters in normal mode.
'properties.kerberos.domain.name' = 'hadoop.System domain name'--This parameter is not required for clusters in normal mode.
);

insert into 
kafka 
select 
* 
from 
stream_mor union all select * from stream_write_cow;

NOTE:

Kafka port
- Value of sasl.port when Authentication Mode of the cluster is Security Mode, 21007 by default.
- Value of port when Authentication Mode of the cluster is Normal Mode, 9092 by default. If the port number is set to 9092, set allow.everyone.if.no.acl.found to true. The procedure is as follows:
  Log in to FusionInsight Manager and choose Cluster > Services > Kafka. On the page that is displayed, click the Configurations tab then the All Configurations sub-tab. On the displayed page, search for allow.everyone.if.no.acl.found, set it to true, and click Save.
System domain name: You can log in to FusionInsight Manager, choose System > Permission > Domain and Mutual Trust, and check the value of Local Domain.

Execute the following script to write data to Kafka. For details, see Managing Messages in Kafka Topics.

sh kafka-console-producer.sh --broker-list IP address of the node where Kafka instances are deployed:Kafka port --topic Topic name --producer.config Client directory/Kafka/kafka/config/producer.properties

In this example, the topic name is writehudi.

sh kafka-console-producer.sh --broker-list IP address of the node where the Kafka instance is deployed:Kafka port ----topic writehudi --producer.config /opt/client/Kafka/kafka/config/producer.properties
Enter the message content.
```
{"uuid": "1","name":"a01","age":10,"ts":10,"p":"1"}
{"uuid": "2","name":"a02","age":20,"ts":20,"p":"2"}
```
Press Enter to send the message.
Consumes Kafka topic data and reads the result of reading the Hudi table from Flink streams.

sh kafka-console-consumer.sh --bootstrap-server IP address of the node where the Kafka Broker instance is deployed:Kafka port --topic Topic name --consumer.config Client directory/Kafka/kafka/config/consumer.properties --from-beginning

In this example, the topic name is readhudi.

sh kafka-console-consumer.sh --bootstrap-server IP address of the node where the Kafka Broker instance is deployed:Kafka port --topic readhudi --consumer.config /opt/client/Kafka/kafka/config/consumer.properties --from-beginning

The read result is as follows (the sequence is not fixed):
```
{"uuid": "1","name":"a01","age":10,"ts":10,"p":"1"}
{"uuid": "2","name":"a02","age":20,"ts":20,"p":"2"}
{"uuid": "1","name":"a01","age":10,"ts":10,"p":"1"}
{"uuid": "2","name":"a02","age":20,"ts":20,"p":"2"}
```

Precautions for Using FlinkSQL Lookup Join Hudi

(This topic is available for MRS 3.5.0 and later.)

The lookup.join.cache.ttl parameter is used to control the loading period of dimension table data. The default value is 60 minutes.
Do not use a Hudi table with over 100,000 rows of records as a dimension table, as the data will be loaded to Flink TaskManager Heap memory.
The new and updated data in the dimension table can be loaded for calculation only after the next loading period.

The following is a SQL example:

CREATE TABLE hudimor(
  uuid VARCHAR(20),
  name VARCHAR(10),
  age INT,
  ts INT,
  `p` VARCHAR(20),
  PRIMARY KEY (uuid) NOT ENFORCED
) PARTITIONED BY (`p`) WITH (
  'connector' = 'hudi',
  'path' = 'hdfs://hacluster/tmp/hudimor',
  'table.type' = 'MERGE_ON_READ',
  'hoodie.datasource.write.recordkey.field' = 'uuid',
  'write.precombine.field' = 'ts',
  'lookup.join.cache.ttl' = '60min'
);
CREATE TABLE datagen(uuid varchar(20), proctime as PROCTIME()) WITH (
  'connector' = 'datagen',
  'rows-per-second' = '1'
);
CREATE TABLE blackhole (
  uuid VARCHAR(20),
  name VARCHAR(10),
  age INT,
  ts INT,
  `p` VARCHAR(20)
) WITH ('connector' = 'blackhole');
insert into
  blackhole
select
  t1.uuid as uuid,
  t2.name as name,
  t2.age as age,
  t2.ts as ts,
  t2.p as p
FROM
  datagen AS t1
  left JOIN hudimor FOR SYSTEM_TIME AS OF t1.proctime AS t2 ON t1.uuid = t2.uuid;

WITH Parameters

**Table 2** WITH parameters
Mode	Configuration Item	Mandatory	Default Value	Description
Read	read.tasks	No	4	Parallelism of the tasks for reading the Hudi table.
	read.streaming.enabled	No	false	Whether to enable stream read.
	read.streaming.start-commit	No	By default, the latest commit is the start position.	Start position (closed interval) of incremental stream and batch consumption in yyyyMMddHHmmss format.
	read.end-commit	No	By default, the latest commit is the end position.	End position (closed interval) of incremental stream and batch consumption in yyyyMMddHHmmss format.
Write	write.tasks	No	4	Parallelism of the tasks for reading data from the Hudi table.
	index.bootstrap.enabled	No	false	Whether to enable index loading. If it is enabled, the latest data in the stored table is loaded to the state at a time. If incremental data needs to be synchronized to full data and there are offline Hoodie tables available, you can enable the index loading function to write data in real time and ensure that data is unique.
	write.index_bootstrap.tasks	No	4	If indexes are loaded slowly when a job is started, you can choose a larger value for this parameter. After that, the efficiency is improved, but checkpoints are blocked in the bootstrap phase.
	compaction.async.enabled	No	true	Whether to enable online compaction
	compaction.schedule.enabled	No	true	Whether to generate a compression plan periodically. You are advised to enable this function even if online compaction is disabled.
	compaction.tasks	No	10	Parallelism of the tasks for compacting data in the Hudi table.
	index.state.ttl	No	7D	Duration for storing indexes. The default value is 7 days. If the value is less than 0, indexes are stored permanently. Indexes are the core data structure for determining whether data is duplicate. For long-time updates, for example, updating data generated one month ago, you need to increase the value of this parameter.

Synchronizing Metadata from Flink On Hudi to Hive

After this feature is enabled, Flink automatically creates a Hudi table on Hive and adds partitions to the Hudi table when writing data to it. Then services such as SparkSQL and Hive can read data form the Hudi table.

The metadata can be synchronized with either of the following methods. The JDBC mode is used as an example in the following steps.

This step is required for MRS 3.2.0 or later.

Synchronizing metadata to Hive in JDBC mode

CREATE TABLE stream_mor(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) PARTITIONED BY (`p`) WITH (
'connector' = 'hudi',
'path' = 'hdfs://hacluster/tmp/hudi/stream_mor',
'table.type' = 'MERGE_ON_READ',
'hive_sync.enable' = 'true',
'hive_sync.table' = 'Name of the table to be synchronized to Hive',
'hive_sync.db' = 'Name of the database to be synchronized to Hive',
'hive_sync.metastore.uris' = 'Value of hive.metastore.uris in the hive-site.xml file on the Hive client',
'hive_sync.jdbc_url' = 'Value of CLIENT_HIVE_URI in the component_env file on the Hive client'
);

NOTICE:

hive_sync.jdbc_url: If the value of CLIENT_HIVE_URI in the Hive client file component_env contains \, delete \.
To use the Hive style partitioning, add the following parameters:
- 'hoodie.datasource.write.hive_style_partitioning' = 'true'
- 'hive_sync.partition_extractor_class' = 'org.apache.hudi.hive.MultiPartKeysValueExtractor'
Flink on Hudi synchronizes data to Hive. Hudi is case sensitive, while Hive is case insensitive. You are not advised to use uppercase letters in fields of Hudi tables. Otherwise, data may fail to be read or written.

Synchronizing metadata to Hive in HMS mode

CREATE TABLE stream_mor(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) PARTITIONED BY (`p`) WITH (
'connector' = 'hudi',
'path' = 'hdfs://hacluster/tmp/hudi/stream_mor',
'table.type' = 'MERGE_ON_READ',
'hive_sync.enable' = 'true',
'hive_sync.table' = 'Name of the table to be synchronized to Hive',
'hive_sync.db' = 'Name of the database to be synchronized to Hive',
'hive_sync.mode' = 'hms',
'hive_sync.metastore.uris' = 'Value of hive.metastore.uris in the hive-site.xml file on the Hive client',
'properties.hive.metastore.kerberos.principal' = 'Value of hive.metastore.kerberos.principal in the hive-site.xml file on the Hive client'
);

Methods with JDBC:

Log in to Manager as user flink_admin and choose Cluster > Services > Flink. In the Basic Information area, click the link on the right of Flink WebUI to access the Flink web UI.

In Basic Parameter, select Enable CheckPoint, set Time Interval(ms) to 60000, and retain the default value for Mode.

CREATE TABLE stream_mor2(
uuid VARCHAR(20),
name VARCHAR(10),
age INT,
ts INT,
`p` VARCHAR(20)
) PARTITIONED BY (`p`) WITH (
'connector' = 'hudi',
'path' = 'hdfs://hacluster/tmp/hudi/stream_mor2',
'table.type' = 'MERGE_ON_READ',
'hoodie.datasource.write.recordkey.field' = 'uuid',
'write.precombine.field' = 'ts',
'write.tasks' = '4',
'hive_sync.enable' = 'true',
'hive_sync.table' = 'Name of the table to be synchronized to Hive, for example, stream_mor2',
'hive_sync.db' = 'Name of the database to be synchronized to Hive, for example, default',
'hive_sync.metastore.uris' = 'Value of hive.metastore.uris in the hive-site.xml file on the Hive client',
'hive_sync.jdbc_url' = 'Value of CLIENT_HIVE_URI in the component_env file on the Hive client'
);
CREATE TABLE datagen (
uuid varchar(20), name varchar(10), age int, ts INT, p varchar(20)
) WITH (
'connector' = 'datagen',
'rows-per-second' = '1',
'fields.p.length' = '1'
);insert into stream_mor2 select * from datagen;

Wait for the Flink job to run for a period of time and continuously write the random test data generated by datagen to the Hudi table. You can click More > Job Monitoring to go to the native UI of Flink and view the job status.
Log in to the node where the client is deployed, load environment variables, run the beeline command to log in to the Hive client, and run SQL statements to check whether the Hudi Sink table is successfully created on Hive and whether data can be read from the table.

cd /opt/hadoopclient

source bigdata_env

beeline

desc formatted default.stream_mor2;

select * from default.stream_mor2 limit 5;

show partitions default.stream_mor2;