Configuring Tez - Amazon EMR
Example configurationTez asynchronous split opening
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

Configuring Tez

You can customize Tez by setting values using the tez-site configuration classification, which configures settings in the tez-site.xml configuration file. For more information, see TezConfiguration in the Apache Tez documentation. To change Hive or Pig to use the Tez execution engine, use the hive-site and pig-properties configuration classifications as appropriate. Examples are shown below.

Example configuration

Example: Customizing the Tez root logging level and setting Tez as the execution engine for Hive and Pig

The example create-cluster command shown below creates a cluster with Tez, Hive, and Pig installed. The command references a file stored in Amazon S3, myConfig.json, which specifies properties for the tez-site classification that sets tez.am.log.level to DEBUG, and sets the execution engine to Tez for Hive and Pig using the hive-site and pig-properties configuration classifications.

Note

Linux line continuation characters (\) are included for readability. They can be removed or used in Linux commands. For Windows, remove them or replace with a caret (^).

aws emr create-cluster --release-label emr-7.0.0 \
--applications Name=Tez Name=Hive Name=Pig --ec2-attributes KeyName=myKey \
--instance-type m5.xlarge --instance-count 3 \
--configurations https://s3.amazonaws.com/mybucket/myfolder/myConfig.json --use-default-roles

Example contents of myConfig.json are shown below.

[
    {
      "Classification": "tez-site",
      "Properties": {
        "tez.am.log.level": "DEBUG"
      }
    },
    {
      "Classification": "hive-site",
      "Properties": {
        "hive.execution.engine": "tez"
      }
    },
    {
      "Classification": "pig-properties",
      "Properties": {
        "exectype": "tez"
      }
    }
  ]
Note

With Amazon EMR version 5.21.0 and later, you can override cluster configurations and specify additional configuration classifications for each instance group in a running cluster. You do this by using the Amazon EMR console, the Amazon Command Line Interface (Amazon CLI), or the Amazon SDK. For more information, see Supplying a Configuration for an Instance Group in a Running Cluster.

Tez asynchronous split opening

When there is a large number of small files in the table path, and a query attempts to read them all, each small file that corresponds to each individual split gets combined under one Tez grouped split. A single mapper then processes the single Tez grouped split. Since the execution is synchronous, each individual split under the grouped split gets processed one by one. This requires RecordReader objects to synchronously process the splits.

Name Classification Description

tez.grouping.split.init.threads

tez-site

Specifies the number of daemon threads that Tez uses to pre-initiate the RecordReaders and open splits. For ACID tables, the maximum supported value of tez.grouping.split.init.threads is 1.

tez.grouping.split.init.recordreaders

tez-site

Specifies the number of RecordReaders to keep pre-initialised by the daemon threads. This can help when Tez grouped split contains a large number of InputSplits. Initialization of RecordReaders to process those input splits can be done asynchronously with daemon threads instead of sequential processing.

Benchmarking for Tez asynchronous split opening

We used the following environments and configurations for benchmarking the Tez asynchronous split opening capability:

  • Benchmark environment – Amazon EMR cluster with 1 primary node that uses m5.16xlarge, and 16 core nodes that use m5.16xlarge.

  • Benchmark configurations – To simulate the scenario for benchmarking where a large number of input splits are in a single Tez grouped split, tez.grouping.split-count is set to 1.

  • Table used for benchmarking – The table contains 200 partitions, with each partition containing a single file. The benchmark is done for when that table contains CSV files, and when that table contains parquet files. Hive query for benchmarking: SELECT COUNT(*) from the table ten times, and take the average runtime.

  • Configurations to enable Tez async split opening – As follows:

    • tez.grouping.split.init.threads = 4

    • tez.grouping.split.init.recordreaders = 10

Dataset Feature disabled (baseline) Feature enabled Improvement

CSV dataset

90.26 seconds

79.20 seconds

12.25%

Parquet dataset

54.67 seconds

42.23 seconds

22.75%