Skip to main content
Version: 3.2

Read data from StarRocks using Spark connector

StarRocks provides a self-developed connector named StarRocks Connector for Apache Spark™ (Spark connector for short) to help you read data from a StarRocks table by using Spark. You can use Spark for complex processing and machine learning on the data you have read from StarRocks.

The Spark connector supports three reading methods: Spark SQL, Spark DataFrame, and Spark RDD.

You can use Spark SQL to create a temporary view on the StarRocks table, and then directly read data from the StarRocks table by using that temporary view.

You can also map the StarRocks table to a Spark DataFrame or a Spark RDD, and then read data from the Spark DataFrame or Spark RDD. We recommend the use of a Spark DataFrame.

NOTICE

Only users with the SELECT privilege on a StarRocks table can read data from this table. You can follow the instructions provided in GRANT to grant the privilege to a user.

Usage notes​

  • You can filter data on StarRocks before you read the data, thereby reducing the amount of data transferred.
  • If the overhead of reading data is substantial, you can employ appropriate table design and filter conditions to prevent Spark from reading an excessive amount of data at a time. As such, you can reduce I/O pressure on your disk and network connection, thereby ensuring routine queries can be run properly.

Version requirements​

Spark connectorSparkStarRocksJavaScala
1.1.23.2, 3.3, 3.4, 3.52.5 and later82.12
1.1.13.2, 3.3, 3.42.5 and later82.12
1.1.03.2, 3.3, 3.42.5 and later82.12
1.0.03.x1.18 and later82.12
1.0.02.x1.18 and later82.11

NOTICE

  • Please see Upgrade Spark connector for behavior changes among different connector versions.
  • The connector does not provide MySQL JDBC driver since version 1.1.1, and you need import the driver to the Spark classpath manually. You can find the driver on Maven Central.
  • In version 1.0.0, the Spark connector only supports reading data from StarRocks. From version 1.1.0 onwards, the Spark connector supports both reading data from and writing data to StarRocks.
  • Version 1.0.0 differs from version 1.1.0 in terms of parameters and data type mappings. See Upgrade Spark connector.
  • In general cases, no new features will be added to version 1.0.0. We recommend that you upgrade your Spark connector at your earliest opportunity.

Obtain Spark connector​

Use one of the following methods to obtain the Spark connector .jar package that suits your business needs:

  • Download a compiled package.
  • Use Maven to add the dependencies required by the Spark connector. (This method is supported only for Spark connector 1.1.0 and later.)
  • Manually compile a package.

Spark connector 1.1.0 and later​

Spark connector .jar packages are named in the following format:

starrocks-spark-connector-${spark_version}_${scala_version}-${connector_version}.jar

For example, if you want to use Spark connector 1.1.0 with Spark 3.2 and Scala 2.12, you can choose starrocks-spark-connector-3.2_2.12-1.1.0.jar.

NOTICE

In normal cases, the latest Spark connector version can be used with the most recent three Spark versions.

Download a compiled package​

You can obtain Spark connector .jar packages of various versions at Maven Central Repository.

Add Maven dependencies​

Configure the dependencies required by the Spark connector as follows:

NOTICE

You must replace spark_version, scala_version, and connector_version with the Spark version, Scala version, and Spark connector version you use.

<dependency>
  <groupId>com.starrocks</groupId>
  <artifactId>starrocks-spark-connector-${spark_version}_${scala_version}</artifactId>
  <version>${connector_version}</version>
</dependency>

For example, if you want to use Spark connector 1.1.0 with Spark 3.2 and Scala 2.12, configure the dependencies as follows:

<dependency>
  <groupId>com.starrocks</groupId>
  <artifactId>starrocks-spark-connector-3.2_2.12</artifactId>
  <version>1.1.0</version>
</dependency>

Manually compile a package​

  1. Download the Spark connector code.

  2. Use the following command to compile the Spark connector:

    NOTICE

    You must replace spark_version with the Spark version you use.

    sh build.sh <spark_version>

    For example, if you want to use the Spark connector with Spark 3.2, compile the Spark connector as follows:

    sh build.sh 3.2

  3. Go to the target/ path, in which a Spark connector .jar package like starrocks-spark-connector-3.2_2.12-1.1.0-SNAPSHOT.jar is generated upon compilation.

    NOTICE

    If you are using a Spark connector version that is not officially released, the name of the generated Spark connector .jar package contains SNAPSHOT as a suffix.

Spark connector 1.0.0​

Download a compiled package​

Manually compile a package​

  1. Download the Spark connector code.

    NOTICE

    You must switch to spark-1.0.

  2. Take one of the following actions to compile the Spark connector:

    • If you are using Spark 2.x, run the following command, which compiles the Spark connector to suit Spark 2.3.4 by default:

      sh build.sh 2

    • If you are using Spark 3.x, run the following command, which compiles the Spark connector to suit Spark 3.1.2 by default:

      sh build.sh 3

  3. Go to the output/ path, in which the starrocks-spark2_2.11-1.0.0.jar file is generated upon compilation. Then, copy the file to the classpath of Spark:

    • If your Spark cluster runs in Local mode, place the file into the jars/ path.
    • If your Spark cluster runs in Yarn mode, place the file into the pre-deployment package.

You can use the Spark connector to read data from StarRocks only after you place the file into the specified location.

Parameters​

This section describes the parameters you need to configure when you use the Spark connector to read data from StarRocks.

Common parameters​

The following parameters apply to all three reading methods: Spark SQL, Spark DataFrame, and Spark RDD.

ParameterDefault valueDescription
starrocks.fenodesNoneThe HTTP URL of the FE in your StarRocks cluster. Format <fe_host>:<fe_http_port>. You can specify multiple URLs, which must be separated by a comma (,).
starrocks.table.identifierNoneThe name of the StarRocks table. Format: <database_name>.<table_name>.
starrocks.request.retries3The maximum number of times that Spark can retry to send a read request o StarRocks.
starrocks.request.connect.timeout.ms30000The maximum amount of time after which a read request sent to StarRocks times out.
starrocks.request.read.timeout.ms30000The maximum amount of time after which the reading for a request sent to StarRocks times out.
starrocks.request.query.timeout.s3600The maximum amount of time after which a query of data from StarRocks times out. The default timeout period is 1 hour. -1 means that no timeout period is specified.
starrocks.request.tablet.sizeInteger.MAX_VALUEThe number of StarRocks tablets grouped into each Spark RDD partition. A smaller value of this parameter indicates that a larger number of Spark RDD partitions will be generated. A larger number of Spark RDD partitions means higher parallelism on Spark but greater pressure on StarRocks.
starrocks.batch.size4096The maximum number of rows that can be read from BEs at a time. Increasing the value of this parameter can reduce the number of connections established between Spark and StarRocks, thereby mitigating extra time overheads caused by network latency.
starrocks.exec.mem.limit2147483648The maximum amount of memory allowed per query. Unit: bytes. The default memory limit is 2 GB.
starrocks.deserialize.arrow.asyncfalseSpecifies whether to support asynchronously converting the Arrow memory format to RowBatches required for the iteration of the Spark connector.
starrocks.deserialize.queue.size64The size of the internal queue that holds tasks for asynchronously converting the Arrow memory format to RowBatches. This parameter is valid when starrocks.deserialize.arrow.async is set to true.
starrocks.filter.queryNoneThe condition based on which you want to filter data on StarRocks. You can specify multiple filter conditions, which must be joined by and. StarRocks filters the data from the StarRocks table based on the specified filter conditions before the data is read by Spark.
starrocks.timezoneDefault timezone of JVMSupported since 1.1.1. The timezone used to convert StarRocks DATETIME to Spark TimestampType. The default is the timezone of JVM returned by ZoneId#systemDefault(). The format could be a timezone name such as Asia/Shanghai, or a zone offset such as +08:00.

Parameters for Spark SQL and Spark DataFrame​

The following parameters apply only to the Spark SQL and Spark DataFrame reading methods.

ParameterDefault valueDescription
starrocks.fe.http.urlNoneThe HTTP IP address of the FE. This parameter is supported from Spark connector 1.1.0 onwards. This parameter is equivalent to starrocks.fenodes. You only need to configure one of them. In Spark connector 1.1.0 and later, we recommend that you use starrocks.fe.http.url because starrocks.fenodes may be deprecated.
starrocks.fe.jdbc.urlNoneThe address that is used to connect to the MySQL server of the FE. Format: jdbc:mysql://<fe_host>:<fe_query_port>.
NOTICE
In Spark connector 1.1.0 and later, this parameter is mandatory.
userNoneThe username of your StarRocks cluster account. The user needs the SELECT privilege on the StarRocks table.
starrocks.userNoneThe username of your StarRocks cluster account. This parameter is supported from Spark connector 1.1.0 onwards. This parameter is equivalent to user. You only need to configure one of them. In Spark connector 1.1.0 and later, we recommend that you use starrocks.user because user may be deprecated.
passwordNoneThe password of your StarRocks cluster account.
starrocks.passwordNoneThe password of your StarRocks cluster account. This parameter is supported from Spark connector 1.1.0 onwards. This parameter is equivalent to password. You only need to configure one of them. In Spark connector 1.1.0 and later, we recommend that you use starrocks.password because password may be deprecated.
starrocks.filter.query.in.max.count100The maximum number of values supported by the IN expression during predicate pushdown. If the number of values specified in the IN expression exceeds this limit, the filter conditions specified in the IN expression are processed on Spark.

Parameters for Spark RDD​

The following parameters apply only to the Spark RDD reading method.

ParameterDefault valueDescription
starrocks.request.auth.userNoneThe username of your StarRocks cluster account.
starrocks.request.auth.passwordNoneThe password of your StarRocks cluster account.
starrocks.read.fieldNoneThe StarRocks table column from which you want to read data. You can specify multiple columns, which must be separated by a comma (,).

Data type mapping between StarRocks and Spark​

Spark connector 1.1.0 and later​

StarRocks data typeSpark data type
BOOLEANDataTypes.BooleanType
TINYINTDataTypes.ByteType
SMALLINTDataTypes.ShortType
INTDataTypes.IntegerType
BIGINTDataTypes.LongType
LARGEINTDataTypes.StringType
FLOATDataTypes.FloatType
DOUBLEDataTypes.DoubleType
DECIMALDecimalType
CHARDataTypes.StringType
VARCHARDataTypes.StringType
STRINGDataTypes.StringType
DATEDataTypes.DateType
DATETIMEDataTypes.TimestampType
JSONDataTypes.StringType
NOTE:
This data type mapping is supported since Spark connector v1.1.2, and requires a StarRocks version of at least 2.5.13, 3.0.3, 3.1.0 or later.
ARRAYUnsupported datatype
HLLUnsupported datatype
BITMAPUnsupported datatype

Spark connector 1.0.0​

StarRocks data typeSpark data type
BOOLEANDataTypes.BooleanType
TINYINTDataTypes.ByteType
SMALLINTDataTypes.ShortType
INTDataTypes.IntegerType
BIGINTDataTypes.LongType
LARGEINTDataTypes.StringType
FLOATDataTypes.FloatType
DOUBLEDataTypes.DoubleType
DECIMALDecimalType
CHARDataTypes.StringType
VARCHARDataTypes.StringType
DATEDataTypes.StringType
DATETIMEDataTypes.StringType
ARRAYUnsupported datatype
HLLUnsupported datatype
BITMAPUnsupported datatype

The processing logic of the underlying storage engine used by StarRocks cannot cover an expected time range when DATE and DATETIME data types are directly used. Therefore, the Spark connector maps the DATE and DATETIME data types from StarRocks to the STRING data type from Spark, and generates readable string texts matching the date and time data read from StarRocks.

Upgrade Spark connector​

Upgrade from version 1.0.0 to version 1.1.0​

  • Since 1.1.1, the Spark connector does not provide mysql-connector-java which is the official JDBC driver for MySQL, because of the limitations of the GPL license used by mysql-connector-java. However, the Spark connector still needs the mysql-connector-java to connect to StarRocks for the table metadata, so you need to add the driver to the Spark classpath manually. You can find the driver on MySQL site or Maven Central.

  • In version 1.1.0, the Spark connector uses JDBC to access StarRocks to obtain more detailed table information. Therefore, you must configure starrocks.fe.jdbc.url.

  • In version 1.1.0, some parameters are renamed. Both the old and new parameters are retained for now. For each pair of equivalent parameters, you only need to configure one of them, but we recommend that you use the new one because the old one may be deprecated.

    • starrocks.fenodes is renamed as starrocks.fe.http.url.
    • user is renamed as starrocks.user.
    • password is renamed as starrocks.password.

  • In version 1.1.0, the mappings of some data types are adjusted based on Spark 3.x:

    • DATE in StarRocks is mapped to DataTypes.DateType (originally DataTypes.StringType) in Spark.
    • DATETIME in StarRocks is mapped to DataTypes.TimestampType (originally DataTypes.StringType) in Spark.

Examples​

The following examples assume you have created a database named test in your StarRocks cluster and you have the permissions of user root. The parameter settings in the examples are based on Spark Connector 1.1.0.

Network configuration​

Ensure that the machine where Spark is located can access the FE nodes of the StarRocks cluster via the http_port (default: 8030) and query_port (default: 9030), and the BE nodes via the be_port (default: 9060).

Data example​

Do as follows to prepare a sample table:

  1. Go to the test database and create a table named score_board.

    MySQL [test]> CREATE TABLE `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`)
    PROPERTIES (
        "replication_num" = "3"
    );

  2. Insert data into the score_board table.

    MySQL [test]> INSERT INTO score_board
    VALUES
        (1, 'Bob', 21),
        (2, 'Stan', 21),
        (3, 'Sam', 22),
        (4, 'Tony', 22),
        (5, 'Alice', 22),
        (6, 'Lucy', 23),
        (7, 'Polly', 23),
        (8, 'Tom', 23),
        (9, 'Rose', 24),
        (10, 'Jerry', 24),
        (11, 'Jason', 24),
        (12, 'Lily', 25),
        (13, 'Stephen', 25),
        (14, 'David', 25),
        (15, 'Eddie', 26),
        (16, 'Kate', 27),
        (17, 'Cathy', 27),
        (18, 'Judy', 27),
        (19, 'Julia', 28),
        (20, 'Robert', 28),
        (21, 'Jack', 29);

  3. Query the score_board table.

    MySQL [test]> SELECT * FROM score_board;
    +------+---------+-------+
    | id   | name    | score |
    +------+---------+-------+
    |    1 | Bob     |    21 |
    |    2 | Stan    |    21 |
    |    3 | Sam     |    22 |
    |    4 | Tony    |    22 |
    |    5 | Alice   |    22 |
    |    6 | Lucy    |    23 |
    |    7 | Polly   |    23 |
    |    8 | Tom     |    23 |
    |    9 | Rose    |    24 |
    |   10 | Jerry   |    24 |
    |   11 | Jason   |    24 |
    |   12 | Lily    |    25 |
    |   13 | Stephen |    25 |
    |   14 | David   |    25 |
    |   15 | Eddie   |    26 |
    |   16 | Kate    |    27 |
    |   17 | Cathy   |    27 |
    |   18 | Judy    |    27 |
    |   19 | Julia   |    28 |
    |   20 | Robert  |    28 |
    |   21 | Jack    |    29 |
    +------+---------+-------+
    21 rows in set (0.01 sec)

Read data using Spark SQL​

  1. Run the following command in the Spark directory to start Spark SQL:

    sh spark-sql

  2. Run the following command to create a temporary view named spark_starrocks on the score_board table which belongs to the test database:

    spark-sql> CREATE TEMPORARY VIEW spark_starrocks
               USING starrocks
               OPTIONS
               (
                   "starrocks.table.identifier" = "test.score_board",
                   "starrocks.fe.http.url" = "<fe_host>:<fe_http_port>",
                   "starrocks.fe.jdbc.url" = "jdbc:mysql://<fe_host>:<fe_query_port>",
                   "starrocks.user" = "root",
                   "starrocks.password" = ""
               );

  3. Run the following command to read data from the temporary view:

    spark-sql> SELECT * FROM spark_starrocks;

    Spark returns the following data:

    1        Bob        21
    2        Stan        21
    3        Sam        22
    4        Tony        22
    5        Alice        22
    6        Lucy        23
    7        Polly        23
    8        Tom        23
    9        Rose        24
    10        Jerry        24
    11        Jason        24
    12        Lily        25
    13        Stephen        25
    14        David        25
    15        Eddie        26
    16        Kate        27
    17        Cathy        27
    18        Judy        27
    19        Julia        28
    20        Robert        28
    21        Jack        29
    Time taken: 1.883 seconds, Fetched 21 row(s)
    22/08/09 15:29:36 INFO thriftserver.SparkSQLCLIDriver: Time taken: 1.883 seconds, Fetched 21 row(s)

Read data using Spark DataFrame​

  1. Run the following command in the Spark directory to start Spark Shell:

    sh spark-shell

  2. Run the following command to create a DataFrame named starrocksSparkDF on the score_board table which belongs to the test database:

    scala> val starrocksSparkDF = spark.read.format("starrocks")
               .option("starrocks.table.identifier", s"test.score_board")
               .option("starrocks.fe.http.url", s"<fe_host>:<fe_http_port>")
               .option("starrocks.fe.jdbc.url", s"jdbc:mysql://<fe_host>:<fe_query_port>")
               .option("starrocks.user", s"root")
               .option("starrocks.password", s"")
               .load()

  3. Read data from the DataFrame. For example, if you want to read the first 10 rows, run the following command:

    scala> starrocksSparkDF.show(10)

    Spark returns the following data:

    +---+-----+-----+
    | id| name|score|
    +---+-----+-----+
    |  1|  Bob|   21|
    |  2| Stan|   21|
    |  3|  Sam|   22|
    |  4| Tony|   22|
    |  5|Alice|   22|
    |  6| Lucy|   23|
    |  7|Polly|   23|
    |  8|  Tom|   23|
    |  9| Rose|   24|
    | 10|Jerry|   24|
    +---+-----+-----+
    only showing top 10 rows

    NOTE

    By default, if you do not specify the number of rows you want to read, Spark returns the first 20 rows.

Read data using Spark RDD​

  1. Run the following command in the Spark directory to start Spark Shell:

    sh spark-shell

  2. Run the following command to create an RDD named starrocksSparkRDD on the score_board table which belongs to the test database.

    scala> import com.starrocks.connector.spark._
    scala> val starrocksSparkRDD = sc.starrocksRDD
               (
               tableIdentifier = Some("test.score_board"),
               cfg = Some(Map(
                   "starrocks.fenodes" -> "<fe_host>:<fe_http_port>",
                   "starrocks.request.auth.user" -> "root",
                   "starrocks.request.auth.password" -> ""
               ))
               )

  3. Read data from the RDD. For example, if you want to read the first 10 elements, run the following command:

    scala> starrocksSparkRDD.take(10)

    Spark returns the following data:

    res0: Array[AnyRef] = Array([1, Bob, 21], [2, Stan, 21], [3, Sam, 22], [4, Tony, 22], [5, Alice, 22], [6, Lucy, 23], [7, Polly, 23], [8, Tom, 23], [9, Rose, 24], [10, Jerry, 24])

    To read the entire RDD, run the following command:

    scala> starrocksSparkRDD.collect()

    Spark returns the following data:

    res1: Array[AnyRef] = Array([1, Bob, 21], [2, Stan, 21], [3, Sam, 22], [4, Tony, 22], [5, Alice, 22], [6, Lucy, 23], [7, Polly, 23], [8, Tom, 23], [9, Rose, 24], [10, Jerry, 24], [11, Jason, 24], [12, Lily, 25], [13, Stephen, 25], [14, David, 25], [15, Eddie, 26], [16, Kate, 27], [17, Cathy, 27], [18, Judy, 27], [19, Julia, 28], [20, Robert, 28], [21, Jack, 29])

Best practices​

When you read data from StarRocks using the Spark connector, you can use the starrocks.filter.query parameter to specify filter conditions based on which Spark prunes partitions, buckets, and prefix indexes to reduce the cost of data pulling. This section uses Spark DataFrame as an example to show how this is achieved.

Environment setup​

ComponentVersion
SparkSpark 2.4.4 and Scala 2.11.12 (OpenJDK 64-Bit Server VM, Java 1.8.0_302)
StarRocks2.2.0
Spark connectorstarrocks-spark2_2.11-1.0.0.jar

Data example​

Do as follows to prepare a sample table:

  1. Go to the test database and create a table named mytable.

    MySQL [test]> CREATE TABLE `mytable`
    (
        `k` int(11) NULL COMMENT "bucket",
        `b` int(11) NULL COMMENT "",
        `dt` datetime NULL COMMENT "",
        `v` int(11) NULL COMMENT ""
    )
    ENGINE=OLAP
    DUPLICATE KEY(`k`,`b`, `dt`)
    COMMENT "OLAP"
    PARTITION BY RANGE(`dt`)
    (
        PARTITION p202201 VALUES [('2022-01-01 00:00:00'), ('2022-02-01 00:00:00')),
        PARTITION p202202 VALUES [('2022-02-01 00:00:00'), ('2022-03-01 00:00:00')),
        PARTITION p202203 VALUES [('2022-03-01 00:00:00'), ('2022-04-01 00:00:00'))
    )
    DISTRIBUTED BY HASH(`k`)
    PROPERTIES (
        "replication_num" = "3"
    );

  2. Insert data into mytable.

    MySQL [test]> INSERT INTO mytable
    VALUES
         (1, 11, '2022-01-02 08:00:00', 111),
         (2, 22, '2022-02-02 08:00:00', 222),
         (3, 33, '2022-03-02 08:00:00', 333);

  3. Query the mytable table.

    MySQL [test]> select * from mytable;
    +------+------+---------------------+------+
    | k    | b    | dt                  | v    |
    +------+------+---------------------+------+
    |    1 |   11 | 2022-01-02 08:00:00 |  111 |
    |    2 |   22 | 2022-02-02 08:00:00 |  222 |
    |    3 |   33 | 2022-03-02 08:00:00 |  333 |
    +------+------+---------------------+------+
    3 rows in set (0.01 sec)

Full table scan​

  1. Run the following command in the Spark directory to create a DataFrame named df on mytable table which belongs to the test database:

    scala>  val df = spark.read.format("starrocks")
            .option("starrocks.table.identifier", s"test.mytable")
            .option("starrocks.fenodes", s"<fe_host>:<fe_http_port>")
            .option("user", s"root")
            .option("password", s"")
            .load()

  2. View the FE log file fe.log of your StarRocks cluster, and find the SQL statement executed to read data. Example:

    2022-08-09 18:57:38,091 INFO (nioEventLoopGroup-3-10|196) [TableQueryPlanAction.executeWithoutPassword():126] receive SQL statement [select `k`,`b`,`dt`,`v` from `test`.`mytable`] from external service [ user ['root'@'%']] for database [test] table [mytable]

  3. In the test database, use EXPLAIN to obtain the execution plan of the SELECT k,b,dt,v from test.mytable statement:

    MySQL [test]> EXPLAIN select `k`,`b`,`dt`,`v` from `test`.`mytable`;
    +-----------------------------------------------------------------------+
    | Explain String                                                        |
    +-----------------------------------------------------------------------+
    | PLAN FRAGMENT 0                                                       |
    |  OUTPUT EXPRS:1: k | 2: b | 3: dt | 4: v                              |
    |   PARTITION: UNPARTITIONED                                            |
    |                                                                       |
    |   RESULT SINK                                                         |
    |                                                                       |
    |   1:EXCHANGE                                                          |
    |                                                                       |
    | PLAN FRAGMENT 1                                                       |
    |  OUTPUT EXPRS:                                                        |
    |   PARTITION: RANDOM                                                   |
    |                                                                       |
    |   STREAM DATA SINK                                                    |
    |     EXCHANGE ID: 01                                                   |
    |     UNPARTITIONED                                                     |
    |                                                                       |
    |   0:OlapScanNode                                                      |
    |      TABLE: mytable                                                   |
    |      PREAGGREGATION: ON                                               |
    |      partitions=3/3                                                   |
    |      rollup: mytable                                                  |
    |      tabletRatio=9/9                                                  |
    |      tabletList=41297,41299,41301,41303,41305,41307,41309,41311,41313 |
    |      cardinality=3                                                    |
    |      avgRowSize=4.0                                                   |
    |      numNodes=0                                                       |
    +-----------------------------------------------------------------------+
    26 rows in set (0.00 sec)

In this example, no pruning is performed. Therefore, Spark scans all of the three partitions (as suggested by partitions=3/3) that hold data, and scans all of the 9 tablets (as suggested by tabletRatio=9/9) in those three partitions.

Partition pruning​

  1. Run the following command, in which you use the starrocks.filter.query parameter to specify a filter condition dt='2022-01-02 08:00:00 for partition pruning, in the Spark directory to create a DataFrame named df on the mytable table which belongs to the test database:

    scala> val df = spark.read.format("starrocks")
           .option("starrocks.table.identifier", s"test.mytable")
           .option("starrocks.fenodes", s"<fe_host>:<fe_http_port>")
           .option("user", s"root")
           .option("password", s"")
           .option("starrocks.filter.query", "dt='2022-01-02 08:00:00'")
           .load()

  2. View the FE log file fe.log of your StarRocks cluster, and find the SQL statement executed to read data. Example:

    2022-08-09 19:02:31,253 INFO (nioEventLoopGroup-3-14|204) [TableQueryPlanAction.executeWithoutPassword():126] receive SQL statement [select `k`,`b`,`dt`,`v` from `test`.`mytable` where dt='2022-01-02 08:00:00'] from external service [ user ['root'@'%']] for database [test] table [mytable]

  3. In the test database, use EXPLAIN to obtain the execution plan of the SELECT k,b,dt,v from test.mytable where dt='2022-01-02 08:00:00' statement:

    MySQL [test]> EXPLAIN select `k`,`b`,`dt`,`v` from `test`.`mytable` where dt='2022-01-02 08:00:00';
    +------------------------------------------------+
    | Explain String                                 |
    +------------------------------------------------+
    | PLAN FRAGMENT 0                                |
    |  OUTPUT EXPRS:1: k | 2: b | 3: dt | 4: v       |
    |   PARTITION: UNPARTITIONED                     |
    |                                                |
    |   RESULT SINK                                  |
    |                                                |
    |   1:EXCHANGE                                   |
    |                                                |
    | PLAN FRAGMENT 1                                |
    |  OUTPUT EXPRS:                                 |
    |   PARTITION: RANDOM                            |
    |                                                |
    |   STREAM DATA SINK                             |
    |     EXCHANGE ID: 01                            |
    |     UNPARTITIONED                              |
    |                                                |
    |   0:OlapScanNode                               |
    |      TABLE: mytable                            |
    |      PREAGGREGATION: ON                        |
    |      PREDICATES: 3: dt = '2022-01-02 08:00:00' |
    |      partitions=1/3                            |
    |      rollup: mytable                           |
    |      tabletRatio=3/3                           |
    |      tabletList=41297,41299,41301              |
    |      cardinality=1                             |
    |      avgRowSize=20.0                           |
    |      numNodes=0                                |
    +------------------------------------------------+
    27 rows in set (0.01 sec)

In this example, only partition pruning is performed, whereas bucket pruning is not. Therefore, Spark scans one of the three partitions (as suggested by partitions=1/3) and all of the tablets (as suggested by tabletRatio=3/3) in that partition.

Bucket pruning​

  1. Run the following command, in which you use the starrocks.filter.query parameter to specify a filter condition k=1 for bucket pruning, in the Spark directory to create a DataFrame named df on the mytable table which belongs to the test database:

    scala> val df = spark.read.format("starrocks")
           .option("starrocks.table.identifier", s"test.mytable")
           .option("starrocks.fenodes", s"<fe_host>:<fe_http_port>")
           .option("user", s"root")
           .option("password", s"")
           .option("starrocks.filter.query", "k=1")
           .load()

  2. View the FE log file fe.log of your StarRocks cluster, and find the SQL statement executed to read data. Example:

    2022-08-09 19:04:44,479 INFO (nioEventLoopGroup-3-16|208) [TableQueryPlanAction.executeWithoutPassword():126] receive SQL statement [select `k`,`b`,`dt`,`v` from `test`.`mytable` where k=1] from external service [ user ['root'@'%']] for database [test] table [mytable]

  3. In the test database, use EXPLAIN to obtain the execution plan of the SELECT k,b,dt,v from test.mytable where k=1 statement:

    MySQL [test]> EXPLAIN select `k`,`b`,`dt`,`v` from `test`.`mytable` where k=1;
    +------------------------------------------+
    | Explain String                           |
    +------------------------------------------+
    | PLAN FRAGMENT 0                          |
    |  OUTPUT EXPRS:1: k | 2: b | 3: dt | 4: v |
    |   PARTITION: UNPARTITIONED               |
    |                                          |
    |   RESULT SINK                            |
    |                                          |
    |   1:EXCHANGE                             |
    |                                          |
    | PLAN FRAGMENT 1                          |
    |  OUTPUT EXPRS:                           |
    |   PARTITION: RANDOM                      |
    |                                          |
    |   STREAM DATA SINK                       |
    |     EXCHANGE ID: 01                      |
    |     UNPARTITIONED                        |
    |                                          |
    |   0:OlapScanNode                         |
    |      TABLE: mytable                      |
    |      PREAGGREGATION: ON                  |
    |      PREDICATES: 1: k = 1                |
    |      partitions=3/3                      |
    |      rollup: mytable                     |
    |      tabletRatio=3/9                     |
    |      tabletList=41299,41305,41311        |
    |      cardinality=1                       |
    |      avgRowSize=20.0                     |
    |      numNodes=0                          |
    +------------------------------------------+
    27 rows in set (0.01 sec)

In this example, only bucket pruning is performed, whereas partition pruning is not. Therefore, Spark scans all of the three partitions (as suggested by partitions=3/3) that hold data, and scans all of the three tablets (as suggested by tabletRatio=3/9) to retrieve Hash values that meet the k = 1 filter condition within those three partitions.

Partition pruning and bucket pruning​

  1. Run the following command, in which you use the starrocks.filter.query parameter to specify two filter conditions k=7 and dt='2022-01-02 08:00:00' for bucket pruning and partition pruning, in the Spark directory to create a DataFrame named df on the mytable table on the test database:

    scala> val df = spark.read.format("starrocks")
           .option("starrocks.table.identifier", s"test.mytable")
           .option("starrocks.fenodes", s"<fe_host>:<fe_http_port>")
           .option("user", s"")
           .option("password", s"")
           .option("starrocks.filter.query", "k=7 and dt='2022-01-02 08:00:00'")
           .load()

  2. View the FE log file fe.log of your StarRocks cluster, and find the SQL statement executed to read data. Example:

    2022-08-09 19:06:34,939 INFO (nioEventLoopGroup-3-18|212) [TableQueryPlanAction.executeWithoutPassword():126] receive SQL statement [select `k`,`b`,`dt`,`v` from `test`.`mytable` where k=7 and dt='2022-01-02 08:00:00'] from external service [ user ['root'@'%']] for database [test] t
    able [mytable]

  3. In the test database, use EXPLAIN to obtain the execution plan of the SELECT k,b,dt,v from test.mytable where k=7 and dt='2022-01-02 08:00:00' statement:

    MySQL [test]> EXPLAIN select `k`,`b`,`dt`,`v` from `test`.`mytable` where k=7 and dt='2022-01-02 08:00:00';
    +----------------------------------------------------------+
    | Explain String                                           |
    +----------------------------------------------------------+
    | PLAN FRAGMENT 0                                          |
    |  OUTPUT EXPRS:1: k | 2: b | 3: dt | 4: v                 |
    |   PARTITION: RANDOM                                      |
    |                                                          |
    |   RESULT SINK                                            |
    |                                                          |
    |   0:OlapScanNode                                         |
    |      TABLE: mytable                                      |
    |      PREAGGREGATION: ON                                  |
    |      PREDICATES: 1: k = 7, 3: dt = '2022-01-02 08:00:00' |
    |      partitions=1/3                                      |
    |      rollup: mytable                                     |
    |      tabletRatio=1/3                                     |
    |      tabletList=41301                                    |
    |      cardinality=1                                       |
    |      avgRowSize=20.0                                     |
    |      numNodes=0                                          |
    +----------------------------------------------------------+
    17 rows in set (0.00 sec)

In this example, both partition pruning and bucket pruning are performed. Therefore, Spark scans only one of the three partitions (as suggested by partitions=1/3) and only one tablet (as suggested by tabletRatio=1/3) in that partition.

Prefix index filtering​

  1. Insert more data records into a partition of the mytable table which belongs to the test database:

    MySQL [test]> INSERT INTO mytable
    VALUES
        (1, 11, "2022-01-02 08:00:00", 111), 
        (3, 33, "2022-01-02 08:00:00", 333), 
        (3, 33, "2022-01-02 08:00:00", 333), 
        (3, 33, "2022-01-02 08:00:00", 333);

  2. Query the mytable table:

    MySQL [test]> SELECT * FROM mytable;
    +------+------+---------------------+------+
    | k    | b    | dt                  | v    |
    +------+------+---------------------+------+
    |    1 |   11 | 2022-01-02 08:00:00 |  111 |
    |    1 |   11 | 2022-01-02 08:00:00 |  111 |
    |    3 |   33 | 2022-01-02 08:00:00 |  333 |
    |    3 |   33 | 2022-01-02 08:00:00 |  333 |
    |    3 |   33 | 2022-01-02 08:00:00 |  333 |
    |    2 |   22 | 2022-02-02 08:00:00 |  222 |
    |    3 |   33 | 2022-03-02 08:00:00 |  333 |
    +------+------+---------------------+------+
    7 rows in set (0.01 sec)

  3. Run the following command, in which you use the starrocks.filter.query parameter to specify a filter condition k=1 for prefix index filtering, in the Spark directory to create a DataFrame named df on the mytable table which belongs to the test database:

    scala> val df = spark.read.format("starrocks")
           .option("starrocks.table.identifier", s"test.mytable")
           .option("starrocks.fenodes", s"<fe_host>:<fe_http_port>")
           .option("user", s"root")
           .option("password", s"")
           .option("starrocks.filter.query", "k=1")
           .load()

  4. In the test database, set is_report_success to true to enable profile reporting:

    MySQL [test]> SET is_report_success = true;
    Query OK, 0 rows affected (0.00 sec)

  5. Use a browser to open the http://<fe_host>:<http_http_port>/query page, and view the profile of the SELECT * FROM mytable where k=1 statement. Example:

    OLAP_SCAN (plan_node_id=0):
      CommonMetrics:
         - CloseTime: 1.255ms
         - OperatorTotalTime: 1.404ms
         - PeakMemoryUsage: 0.00 
         - PullChunkNum: 8
         - PullRowNum: 2
           - __MAX_OF_PullRowNum: 2
           - __MIN_OF_PullRowNum: 0
         - PullTotalTime: 148.60us
         - PushChunkNum: 0
         - PushRowNum: 0
         - PushTotalTime: 0ns
         - SetFinishedTime: 136ns
         - SetFinishingTime: 129ns
      UniqueMetrics:
         - Predicates: 1: k = 1
         - Rollup: mytable
         - Table: mytable
         - BytesRead: 88.00 B
           - __MAX_OF_BytesRead: 88.00 B
           - __MIN_OF_BytesRead: 0.00 
         - CachedPagesNum: 0
         - CompressedBytesRead: 844.00 B
           - __MAX_OF_CompressedBytesRead: 844.00 B
           - __MIN_OF_CompressedBytesRead: 0.00 
         - CreateSegmentIter: 18.582us
         - IOTime: 4.425us
         - LateMaterialize: 17.385us
         - PushdownPredicates: 3
         - RawRowsRead: 2
           - __MAX_OF_RawRowsRead: 2
           - __MIN_OF_RawRowsRead: 0
         - ReadPagesNum: 12
           - __MAX_OF_ReadPagesNum: 12
           - __MIN_OF_ReadPagesNum: 0
         - RowsRead: 2
           - __MAX_OF_RowsRead: 2
           - __MIN_OF_RowsRead: 0
         - ScanTime: 154.367us
         - SegmentInit: 95.903us
           - BitmapIndexFilter: 0ns
           - BitmapIndexFilterRows: 0
           - BloomFilterFilterRows: 0
           - ShortKeyFilterRows: 3
             - __MAX_OF_ShortKeyFilterRows: 3
             - __MIN_OF_ShortKeyFilterRows: 0
           - ZoneMapIndexFilterRows: 0
         - SegmentRead: 2.559us
           - BlockFetch: 2.187us
           - BlockFetchCount: 2
             - __MAX_OF_BlockFetchCount: 2
             - __MIN_OF_BlockFetchCount: 0
           - BlockSeek: 7.789us
           - BlockSeekCount: 2
             - __MAX_OF_BlockSeekCount: 2
             - __MIN_OF_BlockSeekCount: 0
           - ChunkCopy: 25ns
           - DecompressT: 0ns
           - DelVecFilterRows: 0
           - IndexLoad: 0ns
           - PredFilter: 353ns
           - PredFilterRows: 0
           - RowsetsReadCount: 7
           - SegmentsReadCount: 3
             - __MAX_OF_SegmentsReadCount: 2
             - __MIN_OF_SegmentsReadCount: 0
           - TotalColumnsDataPageCount: 8
             - __MAX_OF_TotalColumnsDataPageCount: 8
             - __MIN_OF_TotalColumnsDataPageCount: 0
         - UncompressedBytesRead: 508.00 B
           - __MAX_OF_UncompressedBytesRead: 508.00 B
           - __MIN_OF_UncompressedBytesRead: 0.00

In this example, the filter condition k = 1 can hit the prefix index. Therefore, Spark can filter out three rows (as suggested by ShortKeyFilterRows: 3).