Help Center/ MapReduce Service/ Component Operation Guide (Normal)/ Using Spark2x (for MRS 3.x or Later)/ Spark SQL Enterprise-Class Enhancements/ Configuring Vector-based ORC Data Reading
Updated on 2024-10-08 GMT+08:00
View PDF
Share

Configuring Vector-based ORC Data Reading

Scenario

ORC is a column-based storage format in the Hadoop ecosystem. It originates from Apache Hive and is used to reduce the Hadoop data storage space and accelerate the Hive query speed. Similar to Parquet, ORC is not a pure column-based storage format. In the ORC format, the entire table is split based on the row group, data in each row group is stored by column, and data is compressed as much as possible to reduce storage space consumption. Vector-based ORC data reading significantly improves the ORC data reading performance. In Spark2.3, SparkSQL supports vector-based ORC data reading (this function is supported in earlier Hive versions). Vector-based ORC data reading improves the data reading performance by multiple times.

This feature can be enabled by using the following parameter.
  • spark.sql.orc.enableVectorizedReader: specifies whether vector-based ORC data reading is supported. The default value is true.
  • spark.sql.codegen.wholeStage: specifies whether to compile all stages of multiple operations into a Java method. The default value is true.
  • spark.sql.codegen.maxFields: specifies the maximum number of fields (including nested fields) supported by all stages of codegen. The default value is 100.
  • spark.sql.orc.impl: specifies whether Hive or Spark SQL native is used as the SQL execution engine to read ORC data. The default value is hive.

Parameters

Log in to FusionInsight Manager, choose Cluster > Name of the desired cluster > Services > Spark2x, click the Configurations tab and then All Configurations, and search for the following parameters.

Parameter

Description

Default Value

Value Range

spark.sql.orc.enableVectorizedReader

Specifies whether vector-based ORC data reading is supported. The default value is true.

true

[true,false]

spark.sql.codegen.wholeStage

Specifies whether to compile all stages of multiple operations into a Java method. The default value is true.

true

[true,false]

spark.sql.codegen.maxFields

Specifies the maximum number of fields (including nested fields) supported by all stages of codegen. The default value is 100.

100

Greater than 0

spark.sql.orc.impl

Specifies whether Hive or Spark SQL native is used as the SQL execution engine to read ORC data. The default value is hive.

hive

[hive,native]
  1. To use vector-based ORC data reading of SparkSQL, the following conditions must be met:
    • spark.sql.orc.enableVectorizedReader must be set to true (default value). Generally, the value is not changed.
    • spark.sql.codegen.wholeStage must be set to true (default value). Generally, the value is not changed.
    • The value of spark.sql.codegen.maxFields must be greater than or equal to the number of columns in scheme.
    • All data is of the AtomicType. Specifically, data is not null or of the UDT, array, or map type. If there is data of the preceding types, expected performance cannot be obtained.
    • spark.sql.orc.impl must be set to native. The default value is hive.
  2. If a task is submitted using the client, modification of the following parameters takes effect only after you download the client again: spark.sql.orc.enableVectorizedReader, spark.sql.codegen.wholeStage, spark.sql.codegen.maxFields, and spark.sql.orc.impl.