Continuously load data from Apache Flink®

StarRocks provides a self-developed connector named Flink connector for Apache Flink® (Flink connector for short) to help you load data into a StarRocks table by using Flink. The basic principle is to accumulate the data and then load it all at a time into StarRocks through STREAM LOAD. The Flink connector supports DataStream API, Table API & SQL, and Python API. The Flink connector also has a higher performance than flink-connector-jdbc provided by Apache Flink®.

Version requirements

Obtain Flink connector

You can obtain the Flink connector JAR file in the following ways:

• Directly download the compiled Flink connector JAR file.
• Add the Flink connector as a dependency in your Maven project and then download the JAR file.
• Compile the source code of the Flink connector into a JAR file by yourself.

The naming format of the Flink connector JAR file is as follows:

• Since Flink 1.15, it's flink-connector-starrocks-${connector_version}_flink-${flink_version}.jar. For example, if you install Flink 1.15 and you want to use Flink connector 1.2.7, you can use flink-connector-starrocks-1.2.7_flink-1.15.jar.

• Prior to Flink 1.15, it's flink-connector-starrocks-${connector_version}_flink-${flink_version}_${scala_version}.jar. For example, if you install Flink 1.14 and Scala 2.12 in your environment, and you want to use Flink connector 1.2.7, you can use flink-connector-starrocks-1.2.7_flink-1.14_2.12.jar.

NOTICE

In general, the latest version of the Flink connector only maintains compatibility with the three most recent versions of Flink.

Download the compiled Jar file

Directly download the corresponding version of the Flink connector Jar file from the Maven Central Repository.

Maven Dependency

In your Maven project's pom.xml file, add the Flink connector as a dependency according to the following format. Replace flink_version, scala_version, and connector_version with the respective versions.

• In Flink 1.15 and later

  <dependency>
      <groupId>com.starrocks</groupId>
      <artifactId>flink-connector-starrocks</artifactId>
      <version>${connector_version}_flink-${flink_version}</version>
  </dependency>

• In versions earlier than Flink 1.15

  <dependency>
      <groupId>com.starrocks</groupId>
      <artifactId>flink-connector-starrocks</artifactId>
      <version>${connector_version}_flink-${flink_version}_${scala_version}</version>
  </dependency>

Compile by yourself

1. Download the Flink connector package.

2. Execute the following command to compile the source code of Flink connector into a JAR file. Note that flink_version is replaced with the corresponding Flink version.

   sh build.sh <flink_version>

   For example, if the Flink version in your environment is 1.15, you need to execute the following command:

   sh build.sh 1.15

3. Go to the target/ directory to find the Flink connector JAR file, such as flink-connector-starrocks-1.2.7_flink-1.15-SNAPSHOT.jar, generated upon compilation.

NOTE

The name of Flink connector which is not formally released contains the SNAPSHOT suffix.

Options

Data type mapping between Flink and StarRocks

Usage notes

Exactly Once

• If you want sink to guarantee exactly-once semantics, we recommend you to upgrade StarRocks to 2.5 or later, and Flink connector to 1.2.4 or later

  • Since Flink connector 1.2.4, the exactly-once is redesigned based on Stream Load transaction interface provided by StarRocks since 2.4. Compared to the previous implementation based on non-transactional Stream Load non-transactional interface, the new implementation reduces memory usage and checkpoint overhead, thereby enhancing real-time performance and stability of loading.

  • If the version of StarRocks is earlier than 2.4 or the version of Flink connector is earlier than 1.2.4, the sink will automatically choose the implementation based on Stream Load non-transactional interface.

• Configurations to guarantee exactly-once

  • The value of sink.semantic needs to be exactly-once.

  • If the version of Flink connector is 1.2.8 and later, it is recommended to specify the value of sink.label-prefix. Note that the label prefix must be unique among all types of loading in StarRocks, such as Flink jobs, Routine Load, and Broker Load.

    • If the label prefix is specified, the Flink connector will use the label prefix to clean up lingering transactions that may be generated in some Flink failure scenarios, such as the Flink job fails when a checkpoint is still in progress. These lingering transactions are generally in PREPARED status if you use SHOW PROC '/transactions/<db_id>/running'; to view them in StarRocks. When the Flink job restores from checkpoint, the Flink connector will find these lingering transactions according to the label prefix and some information in checkpoint, and abort them. The Flink connector can not abort them when the Flink job exits because of the two-phase-commit mechanism to implement the exactly-once. When the Flink job exits, the Flink connector has not received the notification from Flink checkpoint coordinator whether the transactions should be included in a successful checkpoint, and it may lead to data loss if these transactions are aborted anyway. You can have an overview about how to achieve end-to-end exactly-once in Flink in this blogpost.

    • If the label prefix is not specified, lingering transactions will be cleaned up by StarRocks only after they time out. However the number of running transactions can reach the limitation of StarRocks max_running_txn_num_per_db if Flink jobs fail frequently before transactions time out. The timeout length is controlled by StarRocks FE configuration prepared_transaction_default_timeout_second whose default value is 86400 (1 day). You can set a smaller value to it to make transactions expired faster when the label prefix is not specified.

• If you are certain that the Flink job will eventually recover from checkpoint or savepoint after a long downtime because of stop or continuous failover, please adjust the following StarRocks configurations accordingly, to avoid data loss.

  • prepared_transaction_default_timeout_second: StarRocks FE configuration, default value is 86400. The value of this configuration needs to be larger than the downtime of the Flink job. Otherwise, the lingering transactions that are included in a successful checkpoint may be aborted because of timeout before you restart the Flink job, which leads to data loss.

    Note that when you set a larger value to this configuration, it is better to specify the value of sink.label-prefix so that the lingering transactions can be cleaned according to the label prefix and some information in checkpoint, instead of due to timeout (which may cause data loss).

  • label_keep_max_second and label_keep_max_num: StarRocks FE configurations, default values are 259200 and 1000 respectively. For details, see FE configurations. The value of label_keep_max_second needs to be larger than the downtime of the Flink job. Otherwise, the Flink connector can not check the state of transactions in StarRocks by using the transaction labels saved in the Flink's savepoint or checkpoint and figure out whether these transactions are committed or not, which may eventually lead to data loss.

  These configurations are mutable and can be modified by using ADMIN SET FRONTEND CONFIG:

    ADMIN SET FRONTEND CONFIG ("prepared_transaction_default_timeout_second" = "3600");
    ADMIN SET FRONTEND CONFIG ("label_keep_max_second" = "259200");
    ADMIN SET FRONTEND CONFIG ("label_keep_max_num" = "1000");

Flush Policy

The Flink connector will buffer the data in memory, and flush them in batch to StarRocks via Stream Load. How the flush is triggered is different between at-least-once and exactly-once.

For at-least-once, the flush will be triggered when any of the following conditions are met:

• the bytes of buffered rows reaches the limit sink.buffer-flush.max-bytes
• the number of buffered rows reaches the limit sink.buffer-flush.max-rows. (Only valid for sink version V1)
• the elapsed time since the last flush reaches the limit sink.buffer-flush.interval-ms
• a checkpoint is triggered

For exactly-once, the flush only happens when a checkpoint is triggered.

Monitoring load metrics

The Flink connector provides the following metrics to monitor loading.

Examples

The following examples show how to use the Flink connector to load data into a StarRocks table with Flink SQL or Flink DataStream.

Preparations

Create a StarRocks table

Create a database test and create a Primary Key table score_board.

CREATE DATABASE `test`;

CREATE TABLE `test`.`score_board`
(
    `id` int(11) NOT NULL COMMENT "",
    `name` varchar(65533) NULL DEFAULT "" COMMENT "",
    `score` int(11) NOT NULL DEFAULT "0" COMMENT ""
)
ENGINE=OLAP
PRIMARY KEY(`id`)
COMMENT "OLAP"
DISTRIBUTED BY HASH(`id`);

Set up Flink environment

• Download Flink binary Flink 1.15.2, and unzip it to directory flink-1.15.2.

• Download Flink connector 1.2.7, and put it into the directory flink-1.15.2/lib.

• Run the following commands to start a Flink cluster:

  cd flink-1.15.2
  ./bin/start-cluster.sh

Run with Flink SQL

• Run the following command to start a Flink SQL client.

  ./bin/sql-client.sh

• Create a Flink table score_board, and insert values into the table via Flink SQL Client.

Note you must define the primary key in the Flink DDL if you want to load data into a Primary Key table of StarRocks. It's optional for other types of StarRocks tables.

```SQL
CREATE TABLE `score_board` (
    `id` INT,
    `name` STRING,
    `score` INT,
    PRIMARY KEY (id) NOT ENFORCED
) WITH (
    'connector' = 'starrocks',
    'jdbc-url' = 'jdbc:mysql://127.0.0.1:9030',
    'load-url' = '127.0.0.1:8030',
    'database-name' = 'test',
    
    'table-name' = 'score_board',
    'username' = 'root',
    'password' = ''
);

INSERT INTO `score_board` VALUES (1, 'starrocks', 100), (2, 'flink', 100);
```

Run with Flink DataStream

There are several ways to implement a Flink DataStream job according to the type of the input records, such as a CSV Java String, a JSON Java String or a custom Java object.

• The input records are CSV-format String. See LoadCsvRecords for a complete example.

  /**
   * Generate CSV-format records. Each record has three values separated by "\t". 
   * These values will be loaded to the columns `id`, `name`, and `score` in the StarRocks table.
   */
  String[] records = new String[]{
          "1\tstarrocks-csv\t100",
          "2\tflink-csv\t100"
  };
  DataStream<String> source = env.fromElements(records);
  
  /**
   * Configure the connector with the required properties.
   * You also need to add properties "sink.properties.format" and "sink.properties.column_separator"
   * to tell the connector the input records are CSV-format, and the column separator is "\t".
   * You can also use other column separators in the CSV-format records,
   * but remember to modify the "sink.properties.column_separator" correspondingly.
   */
  StarRocksSinkOptions options = StarRocksSinkOptions.builder()
          .withProperty("jdbc-url", jdbcUrl)
          .withProperty("load-url", loadUrl)
          .withProperty("database-name", "test")
          .withProperty("table-name", "score_board")
          .withProperty("username", "root")
          .withProperty("password", "")
          .withProperty("sink.properties.format", "csv")
          .withProperty("sink.properties.column_separator", "\t")
          .build();
  // Create the sink with the options.
  SinkFunction<String> starRockSink = StarRocksSink.sink(options);
  source.addSink(starRockSink);

• The input records are JSON-format String. See LoadJsonRecords for a complete example.

  /**
   * Generate JSON-format records. 
   * Each record has three key-value pairs corresponding to the columns `id`, `name`, and `score` in the StarRocks table.
   */
  String[] records = new String[]{
          "{\"id\":1, \"name\":\"starrocks-json\", \"score\":100}",
          "{\"id\":2, \"name\":\"flink-json\", \"score\":100}",
  };
  DataStream<String> source = env.fromElements(records);
  
  /** 
   * Configure the connector with the required properties.
   * You also need to add properties "sink.properties.format" and "sink.properties.strip_outer_array"
   * to tell the connector the input records are JSON-format and to strip the outermost array structure. 
   */
  StarRocksSinkOptions options = StarRocksSinkOptions.builder()
          .withProperty("jdbc-url", jdbcUrl)
          .withProperty("load-url", loadUrl)
          .withProperty("database-name", "test")
          .withProperty("table-name", "score_board")
          .withProperty("username", "root")
          .withProperty("password", "")
          .withProperty("sink.properties.format", "json")
          .withProperty("sink.properties.strip_outer_array", "true")
          .build();
  // Create the sink with the options.
  SinkFunction<String> starRockSink = StarRocksSink.sink(options);
  source.addSink(starRockSink);

• The input records are custom Java objects. See LoadCustomJavaRecords for a complete example.

  • In this example, the input record is a simple POJO RowData.

    public static class RowData {
            public int id;
            public String name;
            public int score;
    
            public RowData() {}
    
            public RowData(int id, String name, int score) {
                this.id = id;
                this.name = name;
                this.score = score;
            }
        }

  • The main program is as follows:

    // Generate records which use RowData as the container.
    RowData[] records = new RowData[]{
            new RowData(1, "starrocks-rowdata", 100),
            new RowData(2, "flink-rowdata", 100),
        };
    DataStream<RowData> source = env.fromElements(records);
    
    // Configure the connector with the required properties.
    StarRocksSinkOptions options = StarRocksSinkOptions.builder()
            .withProperty("jdbc-url", jdbcUrl)
            .withProperty("load-url", loadUrl)
            .withProperty("database-name", "test")
            .withProperty("table-name", "score_board")
            .withProperty("username", "root")
            .withProperty("password", "")
            .build();
    
    /**
     * The connector will use a Java object array (Object[]) to represent a row to be loaded into the StarRocks table,
     * and each element is the value for a column.
     * You need to define the schema of the Object[] which matches that of the StarRocks table.
     */
    TableSchema schema = TableSchema.builder()
            .field("id", DataTypes.INT().notNull())
            .field("name", DataTypes.STRING())
            .field("score", DataTypes.INT())
            // When the StarRocks table is a Primary Key table, you must specify notNull(), for example, DataTypes.INT().notNull(), for the primary key `id`.
            .primaryKey("id")
            .build();
    // Transform the RowData to the Object[] according to the schema.
    RowDataTransformer transformer = new RowDataTransformer();
    // Create the sink with the schema, options, and transformer.
    SinkFunction<RowData> starRockSink = StarRocksSink.sink(schema, options, transformer);
    source.addSink(starRockSink);

  • The RowDataTransformer in the main program is defined as follows:

    private static class RowDataTransformer implements StarRocksSinkRowBuilder<RowData> {
    
        /**
         * Set each element of the object array according to the input RowData.
         * The schema of the array matches that of the StarRocks table.
         */
        @Override
        public void accept(Object[] internalRow, RowData rowData) {
            internalRow[0] = rowData.id;
            internalRow[1] = rowData.name;
            internalRow[2] = rowData.score;
            // When the StarRocks table is a Primary Key table, you need to set the last element to indicate whether the data loading is an UPSERT or DELETE operation.
            internalRow[internalRow.length - 1] = StarRocksSinkOP.UPSERT.ordinal();
        }
    }  

Best practices

Load data to a Primary Key table

This section will show how to load data to a StarRocks Primary Key table to achieve partial updates and conditional updates. You can see Change data through loading for the introduction of those features. These examples use Flink SQL.

Preparations

Create a database test and create a Primary Key table score_board in StarRocks.

CREATE DATABASE `test`;

CREATE TABLE `test`.`score_board`
(
    `id` int(11) NOT NULL COMMENT "",
    `name` varchar(65533) NULL DEFAULT "" COMMENT "",
    `score` int(11) NOT NULL DEFAULT "0" COMMENT ""
)
ENGINE=OLAP
PRIMARY KEY(`id`)
COMMENT "OLAP"
DISTRIBUTED BY HASH(`id`);

Partial update

This example will show how to load data only to columns id and name.

1. Insert two data rows into the StarRocks table score_board in MySQL client.

   mysql> INSERT INTO `score_board` VALUES (1, 'starrocks', 100), (2, 'flink', 100);
   
   mysql> select * from score_board;
   +------+-----------+-------+
   | id   | name      | score |
   +------+-----------+-------+
   |    1 | starrocks |   100 |
   |    2 | flink     |   100 |
   +------+-----------+-------+
   2 rows in set (0.02 sec)

2. Create a Flink table score_board in Flink SQL client.

   • Define the DDL which only includes the columns id and name.

   • Set the option sink.properties.partial_update to true which tells the Flink connector to perform partial updates.

   • If the Flink connector version <= 1.2.7, you also need to set the option sink.properties.columns to id,name,__op to tells the Flink connector which columns need to be updated. Note that you need to append the field __op at the end. The field __op indicates that the data loading is an UPSERT or DELETE operation, and its values are set by the connector automatically.

     CREATE TABLE `score_board` (
         `id` INT,
         `name` STRING,
         PRIMARY KEY (id) NOT ENFORCED
     ) WITH (
         'connector' = 'starrocks',
         'jdbc-url' = 'jdbc:mysql://127.0.0.1:9030',
         'load-url' = '127.0.0.1:8030',
         'database-name' = 'test',
         'table-name' = 'score_board',
         'username' = 'root',
         'password' = '',
         'sink.properties.partial_update' = 'true',
         -- only for Flink connector version <= 1.2.7
         'sink.properties.columns' = 'id,name,__op'
     ); 

3. Insert two data rows into the Flink table. The primary keys of the data rows are as same as these of rows in the StarRocks table. but the values in the column name are modified.

   INSERT INTO `score_board` VALUES (1, 'starrocks-update'), (2, 'flink-update');

4. Query the StarRocks table in MySQL client.

   mysql> select * from score_board;
   +------+------------------+-------+
   | id   | name             | score |
   +------+------------------+-------+
   |    1 | starrocks-update |   100 |
   |    2 | flink-update     |   100 |
   +------+------------------+-------+
   2 rows in set (0.02 sec)

   You can see that only values for name change, and the values for score do not change.

Conditional update

This example will show how to do conditional update according to the value of column score. The update for an id takes effect only when the new value for score is has a greater or equal to the old value.

1. Insert two data rows into the StarRocks table in MySQL client.

   mysql> INSERT INTO `score_board` VALUES (1, 'starrocks', 100), (2, 'flink', 100);
   
   mysql> select * from score_board;
   +------+-----------+-------+
   | id   | name      | score |
   +------+-----------+-------+
   |    1 | starrocks |   100 |
   |    2 | flink     |   100 |
   +------+-----------+-------+
   2 rows in set (0.02 sec)

2. Create a Flink table score_board in the following ways:

   • Define the DDL including all of columns.
   • Set the option sink.properties.merge_condition to score to tell the connector to use the column score as the condition.
   • Set the option sink.version to V1 which tells the connector to use Stream Load.

   CREATE TABLE `score_board` (
       `id` INT,
       `name` STRING,
       `score` INT,
       PRIMARY KEY (id) NOT ENFORCED
   ) WITH (
       'connector' = 'starrocks',
       'jdbc-url' = 'jdbc:mysql://127.0.0.1:9030',
       'load-url' = '127.0.0.1:8030',
       'database-name' = 'test',
       'table-name' = 'score_board',
       'username' = 'root',
       'password' = '',
       'sink.properties.merge_condition' = 'score'
   );

3. Insert two data rows into the Flink table. The primary keys of the data rows are as same as these of rows in the StarRocks table. The first data row has a smaller value in the column score, and the second data row has a larger value in the column score.

   INSERT INTO `score_board` VALUES (1, 'starrocks', 99), (2, 'flink', 101);

4. Query the StarRocks table in MySQL client.

   mysql> select * from score_board;
   +------+-----------+-------+
   | id   | name      | score |
   +------+-----------+-------+
   |    1 | starrocks |   100 |
   |    2 | flink     |   101 |
   +------+-----------+-------+
   2 rows in set (0.02 sec)

You can see that only the value of score in the second data row changes, and the value of score in the first data row does not change.

Load data into columns of BITMAP type

BITMAP is often used to accelerate count distinct, such as counting UV, see Use Bitmap for exact Count Distinct. Here we take the counting of UV as an example to show how to load data into columns of the BITMAP type.

1. Create a StarRocks Aggregate table in MySQL client.

   In the database test, create an Aggregate table page_uv where the column visit_users is defined as the BITMAP type and configured with the aggregate function BITMAP_UNION.

   CREATE TABLE `test`.`page_uv` (
     `page_id` INT NOT NULL COMMENT 'page ID',
     `visit_date` datetime NOT NULL COMMENT 'access time',
     `visit_users` BITMAP BITMAP_UNION NOT NULL COMMENT 'user ID'
   ) ENGINE=OLAP
   AGGREGATE KEY(`page_id`, `visit_date`)
   DISTRIBUTED BY HASH(`page_id`);

2. Create a Flink table in Flink SQL client.

   The column visit_user_id in the Flink table is of BIGINT type, and we want to load this column to the column visit_users of BITMAP type in the StarRocks table. So when defining the DDL of the Flink table, note that:

   • Because Flink does not support BITMAP, you need to define a column visit_user_id as BIGINT type to represent the column visit_users of BITMAP type in the StarRocks table.

   • You need to set the option sink.properties.columns to page_id,visit_date,user_id,visit_users=to_bitmap(visit_user_id), which tells the connector the column mapping beween the Flink table and StarRocks table. Also you need to use to_bitmap function to tell the connector to convert the data of BIGINT type into BITMAP type.

     CREATE TABLE `page_uv` (
         `page_id` INT,
         `visit_date` TIMESTAMP,
         `visit_user_id` BIGINT
     ) WITH (
         'connector' = 'starrocks',
         'jdbc-url' = 'jdbc:mysql://127.0.0.1:9030',
         'load-url' = '127.0.0.1:8030',
         'database-name' = 'test',
         'table-name' = 'page_uv',
         'username' = 'root',
         'password' = '',
         'sink.properties.columns' = 'page_id,visit_date,visit_user_id,visit_users=to_bitmap(visit_user_id)'
     );

3. Load data into Flink table in Flink SQL client.

   INSERT INTO `page_uv` VALUES
      (1, CAST('2020-06-23 01:30:30' AS TIMESTAMP), 13),
      (1, CAST('2020-06-23 01:30:30' AS TIMESTAMP), 23),
      (1, CAST('2020-06-23 01:30:30' AS TIMESTAMP), 33),
      (1, CAST('2020-06-23 02:30:30' AS TIMESTAMP), 13),
      (2, CAST('2020-06-23 01:30:30' AS TIMESTAMP), 23);

4. Calculate page UVs from the StarRocks table in MySQL client.

   MySQL [test]> SELECT `page_id`, COUNT(DISTINCT `visit_users`) FROM `page_uv` GROUP BY `page_id`;
   +---------+-----------------------------+
   | page_id | count(DISTINCT visit_users) |
   +---------+-----------------------------+
   |       2 |                           1 |
   |       1 |                           3 |
   +---------+-----------------------------+
   2 rows in set (0.05 sec)

Load data into columns of HLL type

HLL can be used for approximate count distinct, see Use HLL for approximate count distinct.

Here we take the counting of UV as an example to show how to load data into columns of the HLL type.

1. Create a StarRocks Aggregate table

   In the database test, create an Aggregate table hll_uv where the column visit_users is defined as the HLL type and configured with the aggregate function HLL_UNION.

   CREATE TABLE `hll_uv` (
     `page_id` INT NOT NULL COMMENT 'page ID',
     `visit_date` datetime NOT NULL COMMENT 'access time',
     `visit_users` HLL HLL_UNION NOT NULL COMMENT 'user ID'
   ) ENGINE=OLAP
   AGGREGATE KEY(`page_id`, `visit_date`)
   DISTRIBUTED BY HASH(`page_id`);

2. Create a Flink table in Flink SQL client.

   The column visit_user_id in the Flink table is of BIGINT type, and we want to load this column to the column visit_users of HLL type in the StarRocks table. So when defining the DDL of the Flink table, note that:

   • Because Flink does not support BITMAP, you need to define a column visit_user_id as BIGINT type to represent the column visit_users of HLL type in the StarRocks table.

   • You need to set the option sink.properties.columns to page_id,visit_date,user_id,visit_users=hll_hash(visit_user_id) which tells the connector the column mapping between Flink table and StarRocks table. Also you need to use hll_hash function to tell the connector to convert the data of BIGINT type into HLL type.

     CREATE TABLE `hll_uv` (
         `page_id` INT,
         `visit_date` TIMESTAMP,
         `visit_user_id` BIGINT
     ) WITH (
         'connector' = 'starrocks',
         'jdbc-url' = 'jdbc:mysql://127.0.0.1:9030',
         'load-url' = '127.0.0.1:8030',
         'database-name' = 'test',
         'table-name' = 'hll_uv',
         'username' = 'root',
         'password' = '',
         'sink.properties.columns' = 'page_id,visit_date,visit_user_id,visit_users=hll_hash(visit_user_id)'
     );

3. Load data into Flink table in Flink SQL client.

   INSERT INTO `hll_uv` VALUES
      (3, CAST('2023-07-24 12:00:00' AS TIMESTAMP), 78),
      (4, CAST('2023-07-24 13:20:10' AS TIMESTAMP), 2),
      (3, CAST('2023-07-24 12:30:00' AS TIMESTAMP), 674);

4. Calculate page UVs from the StarRocks table in MySQL client.

   mysql> SELECT `page_id`, COUNT(DISTINCT `visit_users`) FROM `hll_uv` GROUP BY `page_id`;
   **+---------+-----------------------------+
   | page_id | count(DISTINCT visit_users) |
   +---------+-----------------------------+
   |       3 |                           2 |
   |       4 |                           1 |
   +---------+-----------------------------+
   2 rows in set (0.04 sec)