Skip to content

connector-sink.md

Latest commit

 

History

History
896 lines (707 loc) · 70.1 KB

connector-sink.md

File metadata and controls

896 lines (707 loc) · 70.1 KB
Raw

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 that Flink connector accumulates data in memory 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. It has a higher and more stable performance than flink-connector-jdbc provided by Apache Flink®.

NOTICE

Loading data into StarRocks tables with Flink connector needs SELECT and INSERT privileges. If you do not have these privileges, follow the instructions provided in GRANT to grant these privileges to the user that you use to connect to your StarRocks cluster.

Version requirements

Connector Flink StarRocks Java Scala
1.2.15 1.16,1.17,1.18,1.19,1.20 2.1 and later 8 2.11,2.12
1.2.14 1.16,1.17,1.18,1.19,1.20 2.1 and later 8 2.11,2.12
1.2.12 1.16,1.17,1.18,1.19,1.20 2.1 and later 8 2.11,2.12

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 source code.

  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 1.16, you need to execute the following command:

    sh build.sh 1.16

  3. Go to the target/ directory to find the Flink connector JAR file, such as flink-connector-starrocks-1.2.7_flink-1.16-SNAPSHOT.jar, generated upon compilation.

NOTE

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

General Options

Option Required Default value Description
connector Yes NONE The value must be "starrocks".
jdbc-url Yes NONE The address that is used to connect to the MySQL server of the FE. You can specify multiple addresses, which must be separated by a comma (,). Format: jdbc:mysql://<fe_host1>:<fe_query_port1>,<fe_host2>:<fe_query_port2>,<fe_host3>:<fe_query_port3>.
load-url Yes NONE The address that is used to connect to the HTTP server of the FE. You can specify multiple addresses, which must be separated by a semicolon (;). Format: <fe_host1>:<fe_http_port1>;<fe_host2>:<fe_http_port2>.
database-name Yes NONE The name of the StarRocks database into which you want to load data.
table-name Yes NONE The name of the table that you want to use to load data into StarRocks.
username Yes NONE The username of the account that you want to use to load data into StarRocks. The account needs SELECT and INSERT privileges.
password Yes NONE The password of the preceding account.
sink.semantic No at-least-once The semantic guaranteed by sink. Valid values: at-least-once and exactly-once.
sink.version No AUTO The interface used to load data. This parameter is supported from Flink connector version 1.2.4 onwards.
  • V1: Use Stream Load interface to load data. Connectors before 1.2.4 only support this mode.
  • V2: Use Transaction Stream Load interface to load data. It requires StarRocks to be at least version 2.4. Recommends V2 because it optimizes the memory usage and provides a more stable exactly-once implementation.
  • AUTO: If the version of StarRocks supports transaction Stream Load, will choose V2 automatically, otherwise choose V1
sink.label-prefix No NONE The label prefix used by Stream Load. Recommend to configure it if you are using exactly-once with connector 1.2.8 and later. See exactly-once usage notes.
sink.buffer-flush.max-bytes No 94371840(90M) The maximum size of data that can be accumulated in memory before being sent to StarRocks at a time. The maximum value ranges from 64 MB to 10 GB. Setting this parameter to a larger value can improve loading performance but may increase loading latency. This parameter only takes effect when sink.semantic is set to at-least-once. If sink.semantic is set to exactly-once, the data in memory is flushed when a Flink checkpoint is triggered. In this circumstance, this parameter does not take effect.
sink.buffer-flush.max-rows No 500000 The maximum number of rows that can be accumulated in memory before being sent to StarRocks at a time. This parameter is available only when sink.version is V1 and sink.semantic is at-least-once. Valid values: 64000 to 5000000.
sink.buffer-flush.interval-ms No 300000 The interval at which data is flushed. This parameter is available only when sink.semantic is at-least-once. Valid values: 1000 to 3600000. Since 1.2.14, valid values changed to (0, 3600000]. Unit: ms.
sink.max-retries No 3 The number of times that the system retries to perform the Stream Load job. This parameter is available only when you set sink.version to V1. Valid values: 0 to 10.
sink.connect.timeout-ms No 30000 The timeout for establishing HTTP connection. Valid values: 100 to 60000. Unit: ms. Before 1.2.9, the default value is 1000.
sink.socket.timeout-ms No -1 Supported since 1.2.10. The time duration for which the HTTP client waits for data. Unit: ms. The default value -1 means there is no timeout.
sink.sanitize-error-log No false Supported since 1.2.12. When true, sensitive row data and column values in Stream-Load error logs are redacted in both connector and SDK logs. Intended for production security; defaults to false for backward compatibility.
sink.wait-for-continue.timeout-ms No 10000 Supported since 1.2.7. The timeout for waiting response of HTTP 100-continue from the FE. Valid values: 3000 to 600000. Unit: ms
sink.ignore.update-before No true Supported since version 1.2.8. Whether to ignore UPDATE_BEFORE records from Flink when loading data to Primary Key tables. If this parameter is set to false, the record is treated as a delete operation to StarRocks table.
sink.parallelism No NONE The parallelism of loading. Only available for Flink SQL. If this parameter is not specified, Flink planner decides the parallelism. In the scenario of multi-parallelism, users need to guarantee data is written in the correct order.
sink.properties.* No NONE The parameters that control Stream Load behavior. For example, the parameter sink.properties.format specifies the format used for Stream Load, such as CSV or JSON. For a list of supported parameters and their descriptions, see STREAM LOAD.
sink.properties.format No csv The format used for Stream Load. The Flink connector transforms each batch of data to the format before sending them to StarRocks. Valid values: csv and json.
sink.properties.column_separator No \t The column separator for CSV-formatted data.
sink.properties.row_delimiter No \n The row delimiter for CSV-formatted data.
sink.properties.max_filter_ratio No 0 The maximum error tolerance of the Stream Load. It's the maximum percentage of data records that can be filtered out due to inadequate data quality. Valid values: 0 to 1. Default value: 0. See Stream Load for details.
sink.properties.strict_mode No false Specifies whether to enable the strict mode for Stream Load. It affects the loading behavior when there are unqualified rows, such as inconsistent column values. Valid values: true and false. Default value: false. See Stream Load for details.
sink.properties.compression No NONE The compression algorithm used for Stream Load. Valid values: lz4_frame. Compression for json format needs connector 1.2.10 and StarRocks v3.2.7 or later. Compression for csv format needs connector 1.2.11 and there is no requirements for StarRocks version.
sink.properties.prepared_timeout No NONE Supported since 1.2.12 and only effective when sink.version is V2. Requires StarRocks 3.5.4 or later. Sets the timeout in seconds for the Transaction Stream Load phase from PREPARED to COMMITTED. Typically only needed for exactly-once; at-least-once usually does not require setting this (the connector defaults to 300s). If not set in exactly-once, StarRocks FE configuration prepared_transaction_default_timeout_second (default 86400s) applies. See StarRocks Transaction timeout management.
sink.publish-timeout.ms No -1 Supported since 1.2.14 and only effective when sink.version is V2. Timeout in milliseconds for publish phase. If the transaction stays in COMMITTED status longer than this timeout, consider it as success. Default is -1 which means using StarRocks server-side default behavior. When merge commit is enabled, default is 10000ms.

Merge Commit Options

Supported since 1.2.14. Merge commit allows multiple subtasks to merge data into a single Stream Load transaction for better performance. It is enabled by setting sink.properties.enable_merge_commit to true. For more details about the merge commit feature in StarRocks, see Merge Commit parameters.

The following stream load properties are used to control merge commit behavior:

Option Required Default value Description
sink.properties.enable_merge_commit No false Whether to enable merge commit mode.
sink.properties.merge_commit_interval_ms Yes (when merge commit enabled) NONE The merge commit time window in milliseconds. StarRocks merges loading requests received within this window into a single transaction. A larger value improves merging efficiency but increases latency. Must be set when enable_merge_commit is true.
sink.properties.merge_commit_parallel No 3 The degree of parallelism for the loading plan created by StarRocks for each merge commit. This is different from sink.parallelism which controls the parallelism of the Flink sink operator.
sink.properties.merge_commit_async No true The server's return mode for merge commit. The StarRocks default is false (synchronous), but the connector overrides it to true (asynchronous) for better throughput. In async mode, the server returns immediately after receiving the data. The connector leverages Flink's checkpoint mechanism to ensure no data loss under async mode, providing at-least-once guarantee. Users generally do not need to change this to false.
sink.merge-commit.max-concurrent-requests No Integer.MAX_VALUE Max concurrent stream load requests. Set to 0 to ensure in-order (serial) loading, which is useful for Primary Key tables. A negative value is treated as Integer.MAX_VALUE (unlimited concurrency).
sink.merge-commit.chunk.size No 20971520 The maximum size of data (in bytes) accumulated in a chunk before it is flushed and sent to StarRocks via a stream load request. A larger value improves throughput but increases memory usage and latency; a smaller value reduces memory and latency but may lower throughput. When max-concurrent-requests is 0 (in-order mode), default is 500MB because only one request runs at a time so a larger batch maximizes throughput.

Data type mapping between Flink and StarRocks

Flink data type StarRocks data type
BOOLEAN BOOLEAN
TINYINT TINYINT
SMALLINT SMALLINT
INTEGER INTEGER
BIGINT BIGINT
FLOAT FLOAT
DOUBLE DOUBLE
DECIMAL DECIMAL
BINARY INT
CHAR STRING
VARCHAR STRING
STRING STRING
DATE DATE
TIMESTAMP_WITHOUT_TIME_ZONE(N) DATETIME
TIMESTAMP_WITH_LOCAL_TIME_ZONE(N) DATETIME
ARRAY<T> ARRAY<T>
MAP<KT,VT> JSON STRING
ROW<arg T...> JSON STRING

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. You can set a smaller timeout for PREPARED transactions to make them expired faster when the label prefix is not specified. See the following about how to set the prepared timeout.

  • 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.

    • Adjust PREPARED transaction timeout. See the following about how to set the timeout.

      The timeout 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.

  • How to set the timeout for PREPARED transactions

    • For connector 1.2.12+ and StarRocks 3.5.4+, you can set the timeout by configuring the connector parameter sink.properties.prepared_timeout. By default, the value is not set, and it falls back to the StarRocks FE's global configuration prepared_transaction_default_timeout_second (default value is 86400).

    • For other versions of connector or StarRocks, you can set the timeout by configuring the StarRocks FE's global configuration prepared_transaction_default_timeout_second (default value is 86400).

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.

Merge Commit

Merge commit helps scale throughput without proportionally increasing StarRocks transaction overhead. Without merge commit, each Flink sink subtask maintains its own Stream Load transaction, so increasing sink.parallelism leads to more concurrent transactions and higher IO/Compaction costs on StarRocks. Conversely, keeping parallelism low limits the pipeline's overall capacity. With merge commit enabled, data from multiple sink subtasks is merged into a single transaction within each merge window. This allows you to increase sink.parallelism for higher throughput without increasing the number of transactions. For configuration examples, see Load data with merge commit.

Here are some important notes when using merge commit:

  • Single parallelism has no benefit: If the Flink sink parallelism is 1, enabling merge commit provides no benefit since there is only one subtask sending data. It may even introduce additional latency due to the merge commit time window on the StarRocks side.

  • Only at-least-once semantic: Merge commit only provides at-least-once guarantee. It does not support exactly-once semantic. Do not set sink.semantic to exactly-once when merge commit is enabled.

  • Ordering for Primary Key tables: By default, sink.merge-commit.max-concurrent-requests is Integer.MAX_VALUE, which means a single sink subtask may send multiple stream load requests concurrently. This can cause out-of-order loading, which may be problematic for Primary Key tables. To ensure in-order loading, set sink.merge-commit.max-concurrent-requests to 0, but this will reduce throughput. Alternatively, you can use Conditional update to prevent newer data from being overwritten by older data. For configuration examples, see In-order loading for Primary Key tables.

  • End-to-end loading latency: The total loading latency consists of two parts:

    • Connector batching latency: Determined by sink.buffer-flush.interval-ms and sink.merge-commit.chunk.size. Data is flushed from the connector when either the chunk size limit is reached or the flush interval elapses, whichever comes first. The maximum connector-side latency is sink.buffer-flush.interval-ms. A smaller sink.buffer-flush.interval-ms reduces connector-side latency but sends data in smaller batches.
    • StarRocks merge window: Determined by sink.properties.merge_commit_interval_ms. StarRocks waits for this duration to merge requests from multiple subtasks into a single transaction. A larger value improves merging efficiency (more requests merged into one transaction) but increases StarRocks-side latency.
    • As a general guideline, set sink.buffer-flush.interval-ms to be smaller than or equal to sink.properties.merge_commit_interval_ms, so that each subtask can flush at least once within each merge window. For example, if merge_commit_interval_ms is 10000 (10s), you could set sink.buffer-flush.interval-ms to 5000 (5s) or less.

  • Tuning sink.parallelism and sink.properties.merge_commit_parallel: These two parameters control parallelism at different layers and should be tuned independently:

    • sink.parallelism controls the number of Flink sink subtasks. Each subtask buffers and sends data to StarRocks. Increase this value when Flink sink operators are CPU- or memory-bound — you can monitor Flink's per-operator CPU and memory usage to determine whether more subtasks are needed.
    • sink.properties.merge_commit_parallel controls the degree of parallelism for the loading plan that StarRocks creates for each merge commit transaction. Increase this value when StarRocks becomes the bottleneck. You can monitor the StarRocks metrics merge_commit_pending_total (number of pending merge commit tasks) and merge_commit_pending_bytes (bytes held by pending tasks) to determine whether more parallelism is needed on the StarRocks side — sustained high values indicate that the loading plan cannot keep up with incoming data.

  • Relationship between sink.merge-commit.chunk.size and sink.buffer-flush.max-bytes:

    • sink.merge-commit.chunk.size controls the maximum data size per individual stream load request (per chunk). When data in a chunk reaches this size, it is flushed immediately.
    • sink.buffer-flush.max-bytes controls the total memory limit for all cached data across all tables. When the total cached data exceeds this limit, the connector will evict chunks early to free memory.
    • Therefore, sink.buffer-flush.max-bytes should be set larger than sink.merge-commit.chunk.size to allow at least one full chunk to be accumulated. In general, sink.buffer-flush.max-bytes should be several times larger than sink.merge-commit.chunk.size, especially when there are multiple tables or high concurrency.

Monitoring load metrics

The Flink connector provides the following metrics to monitor loading.

Name Type Description
totalFlushBytes counter successfully flushed bytes.
totalFlushRows counter number of rows successfully flushed.
totalFlushSucceededTimes counter number of times that the data-batch been successfully flushed.
totalFlushFailedTimes counter number of times that the flushing been failed.
totalFilteredRows counter number of rows filtered which is also included in totalFlushRows.

Flink CDC Synchronization (Schema Change Support)

You can use Flink CDC framework to easily build a streaming ELT pipeline from CDC sources (such as MySQL, Kafka) to StarRocks. The pipeline can synchronize whole database, merged sharding tables, and schema changes from sources to StarRocks.

The flink connector has been integrated into the framework as StarRocks Pipeline Connector. The pipeline connector supports

  • create database/table automatically
  • schema change synchronization
  • full and incremental data synchronization

You can see Streaming ELT from MySQL to StarRocks using Flink CDC for a quick start.

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.

    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 Flink connector with the required properties.
 * You also need to add properties "sink.properties.format" and "sink.properties.column_separator"
 * to tell the Flink 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 Flink connector with the required properties.
 * You also need to add properties "sink.properties.format" and "sink.properties.strip_outer_array"
 * to tell the Flink 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 Flink 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 Flink 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 Flink 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 Flink connector to use the column score as the condition.
    • Set the option sink.version to V1 or V2. Both support conditional update.
    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',
    'sink.version' = 'V1'
      );

  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-update', 99), (2, 'flink-update', 101);

  4. Query the StarRocks table in MySQL client.

    mysql> select * from score_board;
+------+--------------+-------+
| id   | name         | score |
+------+--------------+-------+
|    1 | starrocks    |   100 |
|    2 | flink-update |   101 |
+------+--------------+-------+
2 rows in set (0.03 sec)

You can see that only the values of the second data row change, and the values of the first data row do not change.

Load data with merge commit

This section shows how to use merge commit to improve loading throughput when you have multiple Flink sink subtasks writing to the same StarRocks table. These examples use Flink SQL and assume StarRocks v3.4.0 or later.

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`);

Basic configuration

This Flink SQL enables merge commit with a 10-second merge window. Data from all sink subtasks is merged into a single transaction within each window.

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.enable_merge_commit' = 'true',
    'sink.properties.merge_commit_interval_ms' = '10000',
    'sink.buffer-flush.interval-ms' = '5000'
);

Insert data into the Flink table. The data will be loaded into StarRocks via merge commit.

INSERT INTO `score_board` VALUES (1, 'starrocks', 100), (2, 'flink', 95), (3, 'spark', 90);

In-order loading for Primary Key tables

By default, a single sink subtask may send multiple stream load requests concurrently, which can cause out-of-order loading. For Primary Key tables where data ordering matters, there are two approaches to handle this.

Approach 1: Use sink.merge-commit.max-concurrent-requests

Set sink.merge-commit.max-concurrent-requests to 0 to ensure each subtask sends requests one at a time. This guarantees in-order loading but may reduce throughput.

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.enable_merge_commit' = 'true',
    'sink.properties.merge_commit_interval_ms' = '10000',
    'sink.buffer-flush.interval-ms' = '5000',
    'sink.merge-commit.max-concurrent-requests' = '0'
);

INSERT INTO `score_board` VALUES (1, 'starrocks', 100), (2, 'flink', 95), (3, 'spark', 90);

Approach 2: Use conditional update

If you want to keep concurrent requests for higher throughput but still prevent older data from overwriting newer data, you can use Conditional update. Set sink.properties.merge_condition to a column (for example, a version or timestamp column) so that an update only takes effect when the incoming value is greater than or equal to the existing value.

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.enable_merge_commit' = 'true',
    'sink.properties.merge_commit_interval_ms' = '10000',
    'sink.buffer-flush.interval-ms' = '5000',
    'sink.properties.merge_condition' = 'score'
);

INSERT INTO `score_board` VALUES (1, 'starrocks', 100), (2, 'flink', 95), (3, 'spark', 90);

With this configuration, concurrent requests are allowed (default sink.merge-commit.max-concurrent-requests is Integer.MAX_VALUE), but an update to a row only takes effect when the new score is greater than or equal to the existing score. This prevents newer data from being overwritten by older data even under out-of-order loading.

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 Flink connector the column mapping beween the Flink table and StarRocks table. Also you need to use to_bitmap function to tell the Flink 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 Flink connector the column mapping between Flink table and StarRocks table. Also you need to use hll_hash function to tell the Flink 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)