Spark Doris Connector

Spark Doris Connector can support reading data stored in Doris and writing data to Doris through Spark.

Github: https://github.com/apache/doris-spark-connector

  • Support reading data from Doris.
  • Support Spark DataFrame batch/stream writing data to Doris
  • You can map the Doris table to DataFrame or RDD, it is recommended to use DataFrame.
  • Support the completion of data filtering on the Doris side to reduce the amount of data transmission.

Version Compatibility

ConnectorSparkDorisJavaScala
1.3.13.4 ~ 3.1, 2.4, 2.31.0 +82.12, 2.11
1.2.03.2, 3.1, 2.31.0 +82.12, 2.11
1.1.03.2, 3.1, 2.31.0 +82.12, 2.11
1.0.13.1, 2.30.12 - 0.1582.12, 2.11

Build and Install

Ready to work

  1. Modify the custom_env.sh.tpl file and rename it to custom_env.sh

  2. Execute following command in source dir: sh build.sh Follow the prompts to enter the Scala and Spark versions you need to start compiling.

After the compilation is successful, the target jar package will be generated in the dist directory, such as: spark-doris-connector-3.1_2.12-1.2.0-SNAPSHOT.jar. Copy this file to ClassPath in Spark to use Spark-Doris-Connector. For example, Spark running in Local mode, put this file in the jars/ folder. Spark running in Yarn cluster mode, put this file in the pre-deployment package.

For example upload spark-doris-connector-3.1_2.12-1.2.0-SNAPSHOT.jar to hdfs and add hdfs file path in spark.yarn.jars.

  1. Upload spark-doris-connector-3.1_2.12-1.2.0-SNAPSHOT.jar Jar to hdfs.
  1. hdfs dfs -mkdir /spark-jars/
  2. hdfs dfs -put /your_local_path/spark-doris-connector-3.1_2.12-1.2.0-SNAPSHOT.jar /spark-jars/
  1. Add spark-doris-connector-3.1_2.12-1.2.0-SNAPSHOT.jar dependency in Cluster.
  1. spark.yarn.jars=hdfs:///spark-jars/spark-doris-connector-3.1_2.12-1.2.0-SNAPSHOT.jar

Using Maven

  1. <dependency>
  2. <groupId>org.apache.doris</groupId>
  3. <artifactId>spark-doris-connector-3.4_2.12</artifactId>
  4. <version>1.3.0</version>
  5. </dependency>

Notes

Please replace the Connector version according to the different Spark and Scala versions.

Example

Read

SQL

  1. CREATE
  2. TEMPORARY VIEW spark_doris
  3. USING doris
  4. OPTIONS(
  5. "table.identifier"="$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME",
  6. "fenodes"="$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT",
  7. "user"="$YOUR_DORIS_USERNAME",
  8. "password"="$YOUR_DORIS_PASSWORD"
  9. );
  10. SELECT *
  11. FROM spark_doris;

DataFrame

  1. val dorisSparkDF = spark.read.format("doris")
  2. .option("doris.table.identifier", "$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME")
  3. .option("doris.fenodes", "$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT")
  4. .option("user", "$YOUR_DORIS_USERNAME")
  5. .option("password", "$YOUR_DORIS_PASSWORD")
  6. .load()
  7. dorisSparkDF.show(5)

RDD

  1. import org.apache.doris.spark._
  2. val dorisSparkRDD = sc.dorisRDD(
  3. tableIdentifier = Some("$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME"),
  4. cfg = Some(Map(
  5. "doris.fenodes" -> "$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT",
  6. "doris.request.auth.user" -> "$YOUR_DORIS_USERNAME",
  7. "doris.request.auth.password" -> "$YOUR_DORIS_PASSWORD"
  8. ))
  9. )
  10. dorisSparkRDD.collect()

pySpark

  1. dorisSparkDF = spark.read.format("doris")
  2. .option("doris.table.identifier", "$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME")
  3. .option("doris.fenodes", "$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT")
  4. .option("user", "$YOUR_DORIS_USERNAME")
  5. .option("password", "$YOUR_DORIS_PASSWORD")
  6. .load()
  7. // show 5 lines data
  8. dorisSparkDF.show(5)

Write

SQL

  1. CREATE
  2. TEMPORARY VIEW spark_doris
  3. USING doris
  4. OPTIONS(
  5. "table.identifier"="$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME",
  6. "fenodes"="$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT",
  7. "user"="$YOUR_DORIS_USERNAME",
  8. "password"="$YOUR_DORIS_PASSWORD"
  9. );
  10. INSERT INTO spark_doris
  11. VALUES ("VALUE1", "VALUE2", ...);
  12. # or
  13. INSERT INTO spark_doris
  14. SELECT *
  15. FROM YOUR_TABLE
  16. # or
  17. INSERT OVERWRITE
  18. SELECT *
  19. FROM YOUR_TABLE

DataFrame(batch/stream)

  1. ## batch sink
  2. val mockDataDF = List(
  3. (3, "440403001005", "21.cn"),
  4. (1, "4404030013005", "22.cn"),
  5. (33, null, "23.cn")
  6. ).toDF("id", "mi_code", "mi_name")
  7. mockDataDF.show(5)
  8. mockDataDF.write.format("doris")
  9. .option("doris.table.identifier", "$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME")
  10. .option("doris.fenodes", "$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT")
  11. .option("user", "$YOUR_DORIS_USERNAME")
  12. .option("password", "$YOUR_DORIS_PASSWORD")
  13. //other options
  14. //specify the fields to write
  15. .option("doris.write.fields", "$YOUR_FIELDS_TO_WRITE")
  16. // Support setting Overwrite mode to overwrite data
  17. // .option("save_mode", SaveMode.Overwrite)
  18. .save()
  19. ## stream sink(StructuredStreaming)
  20. ### Result DataFrame with structured data, the configuration method is the same as the batch mode.
  21. val sourceDf = spark.readStream.
  22. .format("your_own_stream_source")
  23. .load()
  24. val resultDf = sourceDf.<transformations>
  25. resultDf.writeStream
  26. .format("doris")
  27. .option("checkpointLocation", "$YOUR_CHECKPOINT_LOCATION")
  28. .option("doris.table.identifier", "$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME")
  29. .option("doris.fenodes", "$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT")
  30. .option("user", "$YOUR_DORIS_USERNAME")
  31. .option("password", "$YOUR_DORIS_PASSWORD")
  32. .start()
  33. .awaitTermination()
  34. ### There is a column value in the Result DataFrame that can be written directly, such as the value in the kafka message that conforms to the import format
  35. val kafkaSource = spark.readStream
  36. .format("kafka")
  37. .option("kafka.bootstrap.servers", "$YOUR_KAFKA_SERVERS")
  38. .option("startingOffsets", "latest")
  39. .option("subscribe", "$YOUR_KAFKA_TOPICS")
  40. .load()
  41. kafkaSource.selectExpr("CAST(key AS STRING)", "CAST(value as STRING)")
  42. .writeStream
  43. .format("doris")
  44. .option("checkpointLocation", "$YOUR_CHECKPOINT_LOCATION")
  45. .option("doris.table.identifier", "$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME")
  46. .option("doris.fenodes", "$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT")
  47. .option("user", "$YOUR_DORIS_USERNAME")
  48. .option("password", "$YOUR_DORIS_PASSWORD")
  49. // Set this option to true, and the value column in the Kafka message will be written directly without processing.
  50. .option("doris.sink.streaming.passthrough", "true")
  51. .option("doris.sink.properties.format", "json")
  52. //other options
  53. .start()
  54. .awaitTermination()

Configuration

General

KeyDefault ValueComment
doris.fenodesDoris FE http address, support multiple addresses, separated by commas
doris.table.identifierDoris table identifier, eg, db1.tbl1
doris.request.retries3Number of retries to send requests to Doris
doris.request.connect.timeout.ms30000Connection timeout for sending requests to Doris
doris.request.read.timeout.ms30000Read timeout for sending request to Doris
doris.request.query.timeout.s3600Query the timeout time of doris, the default is 1 hour, -1 means no timeout limit
doris.request.tablet.sizeInteger.MAX_VALUEThe number of Doris Tablets corresponding to an RDD Partition. The smaller this value is set, the more partitions will be generated. This will increase the parallelism on the Spark side, but at the same time will cause greater pressure on Doris.
doris.read.fieldList of column names in the Doris table, separated by commas
doris.batch.size1024The maximum number of rows to read data from BE at one time. Increasing this value can reduce the number of connections between Spark and Doris. Thereby reducing the extra time overhead caused by network delay.
doris.exec.mem.limit2147483648Memory limit for a single query. The default is 2GB, in bytes.
doris.deserialize.arrow.asyncfalseWhether to support asynchronous conversion of Arrow format to RowBatch required for spark-doris-connector iteration
doris.deserialize.queue.size64Asynchronous conversion of the internal processing queue in Arrow format takes effect when doris.deserialize.arrow.async is true
doris.write.fieldsSpecifies the fields (or the order of the fields) to write to the Doris table, fileds separated by commas.
By default, all fields are written in the order of Doris table fields.
doris.sink.batch.size100000Maximum number of lines in a single write BE
doris.sink.max-retries0Number of retries after writing BE failed
doris.sink.properties.formatData format of the stream load.
Supported formats: csv, json, arrow(since version 1.4.0)
More Multi-parameter details
doris.sink.properties.*Import parameters for Stream Load.
For example:
Specify column separator: ‘doris.sink.properties.column_separator’ = ‘,’.
More parameter details
doris.sink.task.partition.sizeThe number of partitions corresponding to the Writing task. After filtering and other operations, the number of partitions written in Spark RDD may be large, but the number of records corresponding to each Partition is relatively small, resulting in increased writing frequency and waste of computing resources. The smaller this value is set, the less Doris write frequency and less Doris merge pressure. It is generally used with doris.sink.task.use.repartition.
doris.sink.task.use.repartitionfalseWhether to use repartition mode to control the number of partitions written by Doris. The default value is false, and coalesce is used (note: if there is no Spark action before the write, the whole computation will be less parallel). If it is set to true, then repartition is used (note: you can set the final number of partitions at the cost of shuffle).
doris.sink.batch.interval.ms50The interval time of each batch sink, unit ms.
doris.sink.enable-2pcfalseWhether to enable two-stage commit. When enabled, transactions will be committed at the end of the job, and all pre-commit transactions will be rolled back when some tasks fail.
doris.sink.auto-redirecttrueWhether to redirect StreamLoad requests. After being turned on, StreamLoad will write through FE and no longer obtain BE information explicitly.

SQL & Dataframe Configuration

KeyDefault ValueComment
userDoris username
passwordDoris password
doris.filter.query.in.max.count100In the predicate pushdown, the maximum number of elements in the in expression value list. If this number is exceeded, the in-expression conditional filtering is processed on the Spark side.
doris.ignore-typeIn a temporary view, specify the field types to ignore when reading the schema.
eg: when ‘doris.ignore-type’=’bitmap,hll’

Structured Streaming Configuration

KeyDefault ValueComment
doris.sink.streaming.passthroughfalseWrite the value of the first column directly without processing.

RDD Configuration

KeyDefault ValueComment
doris.request.auth.userDoris username
doris.request.auth.passwordDoris password
doris.filter.queryFilter expression of the query, which is transparently transmitted to Doris. Doris uses this expression to complete source-side data filtering.

Spark Doris Connector - 图1tip

  1. In Spark SQL, when writing data through insert into, if the target table of doris contains BITMAP or HLL type data, you need to set the parameter doris.ignore-type to the corresponding type, and set doris.write.fields maps the corresponding columns, the usage is as follows:

    BITMAP

    1. CREATE TEMPORARY VIEW spark_doris
    2. USING doris
    3. OPTIONS(
    4. "table.identifier"="$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME",
    5. "fenodes"="$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT",
    6. "user"="$YOUR_DORIS_USERNAME",
    7. "password"="$YOUR_DORIS_PASSWORD"
    8. "doris.ignore-type"="bitmap",
    9. "doris.write.fields"="col1,col2,col3,bitmap_col2=to_bitmap(col2),bitmap_col3=bitmap_hash(col3)"
    10. );

    HLL

    1. CREATE TEMPORARY VIEW spark_doris
    2. USING doris
    3. OPTIONS(
    4. "table.identifier"="$YOUR_DORIS_DATABASE_NAME.$YOUR_DORIS_TABLE_NAME",
    5. "fenodes"="$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT",
    6. "user"="$YOUR_DORIS_USERNAME",
    7. "password"="$YOUR_DORIS_PASSWORD"
    8. "doris.ignore-type"="hll",
    9. "doris.write.fields"="col1,hll_col1=hll_hash(col1)"
    10. );
  2. Since version 1.3.0, the default value of doris.sink.max-retries configuration is 0, which means no retries are performed by default. When this parameter is set greater than 0, batch-level failure retries will be performed, and data of the configured size of doris.sink.batch.size will be cached in the Spark Executor memory. The memory allocation may need to be appropriately increased.

  3. Since version 1.3.0, overwrite mode insertion is supported (only full table-level overwrite insertion is supported). The specific usage is as follows

    DataFrame

    1. resultDf.format("doris")
    2. .option("doris.fenodes","$YOUR_DORIS_FE_HOSTNAME:$YOUR_DORIS_FE_RESFUL_PORT")
    3. // your own options
    4. .option("save_mode", SaveMode.Overwrite)
    5. .save()

    SQL

    1. INSERT OVERWRITE your_target_table
    2. SELECT * FROM your_source_table

Doris & Spark Column Type Mapping

Doris TypeSpark Type
NULL_TYPEDataTypes.NullType
BOOLEANDataTypes.BooleanType
TINYINTDataTypes.ByteType
SMALLINTDataTypes.ShortType
INTDataTypes.IntegerType
BIGINTDataTypes.LongType
FLOATDataTypes.FloatType
DOUBLEDataTypes.DoubleType
DATEDataTypes.DateType
DATETIMEDataTypes.StringType1
DECIMALDecimalType
CHARDataTypes.StringType
LARGEINTDecimalType
VARCHARDataTypes.StringType
TIMEDataTypes.DoubleType
HLLUnsupported datatype
BitmapUnsupported datatype
  • Note: In Connector, DATETIME is mapped to String. Due to the processing logic of the Doris underlying storage engine, when the time type is used directly, the time range covered cannot meet the demand. So use String type to directly return the corresponding time readable text.