This page describes in detail how to run end to end tests on a hudi dataset that helps in improving our confidence in a release as well as perform large scale performance benchmarks.
hdfs, parquet, spark, hive and avro.payload size, number of updates, number of inserts, number of partitionsinsert, bulk_insert, upsert, compact, queryThe Hudi test suite runs as a long running spark job. The suite is divided into the following high level components :
This component does the work of generating the workload; inserts, upserts etc.
Depending on the type of workload generated, data is either ingested into the target hudi dataset or the corresponding workload operation is executed. For example compaction does not necessarily need a workload to be generated/ingested but can require an execution.
The test suite supports different types of operations besides ingestion such as Hive Query execution, Clean action etc.
org.apache.hudi.integ.testsuite.HoodieTestSuiteJob.java - Entry Point of the hudi test suite job. This class wraps all the functionalities required to run a configurable integration suite.
org.apache.hudi.integ.testsuite.HoodieTestSuiteJob.HoodieTestSuiteConfig - Config class that drives the behavior of the integration test suite. This class extends from com.uber.hoodie.utilities.DeltaStreamerConfig. Look at link#HudiDeltaStreamer page to learn about all the available configs applicable to your test suite.
There are 2 ways to generate a workload pattern
1.Programmatically
You can create a DAG of operations programmatically - take a look at WorkflowDagGenerator class. Once you're ready with the DAG you want to execute, simply pass the class name as follows:
spark-submit ... ... --class org.apache.hudi.integ.testsuite.HoodieTestSuiteJob --workload-generator-classname org.apache.hudi.integ.testsuite.dag.scheduler.<your_workflowdaggenerator> ...
2.YAML file
Choose to write up the entire DAG of operations in YAML, take a look at complex-dag-cow.yaml or complex-dag-mor.yaml. Once you're ready with the DAG you want to execute, simply pass the yaml file path as follows:
spark-submit ... ... --class org.apache.hudi.integ.testsuite.HoodieTestSuiteJob --workload-yaml-path /path/to/your-workflow-dag.yaml ...
The test suite can be found in the hudi-integ-test module. Use the prepare_integration_suite.sh script to build the test suite, you can provide different parameters to the script.
shell$ ./prepare_integration_suite.sh --help Usage: prepare_integration_suite.sh --spark-command, prints the spark command -h, hdfs-version -s, spark version -p, parquet version -a, avro version -s, hive version
shell$ ./prepare_integration_suite.sh .... .... Final command : mvn clean install -DskipTests
Copy over the necessary files and jars that are required to your cluster and then run the following spark-submit command after replacing the correct values for the parameters. NOTE : The properties-file should have all the necessary information required to ingest into a Hudi dataset. For more information on what properties need to be set, take a look at the test suite section under demo steps.
shell$ ./prepare_integration_suite.sh --spark-command spark-submit --packages com.databricks:spark-avro_2.11:4.0.0 --master prepare_integration_suite.sh --deploy-mode --properties-file --class org.apache.hudi.integ.testsuite.HoodieTestSuiteJob target/hudi-integ-test-0.6 .0-SNAPSHOT.jar --source-class --source-ordering-field --input-base-path --target-base-path --target-table --props --storage-type --payload-class --workload-yaml-path --input-file-size --<deltastreamer-ingest>
Take a look at the TestHoodieTestSuiteJob to check how you can run the entire suite using JUnit.
Start the Hudi Docker demo:
docker/setup_demo.sh
NOTE: We need to make a couple of environment changes for Hive 2.x support. This will be fixed once Hudi moves to Spark 3.x. Execute below if you are using Hudi query node in your dag. If not, below section is not required. Also, for longer running tests, go to next section.
docker exec -it adhoc-2 bash cd /opt/spark/jars rm /opt/spark/jars/hive* rm spark-hive-thriftserver_2.11-2.4.4.jar wget https://repo1.maven.org/maven2/org/apache/spark/spark-hive-thriftserver_2.12/3.0.0-preview2/spark-hive-thriftserver_2.12-3.0.0-preview2.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-common/2.3.1/hive-common-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-exec/2.3.1/hive-exec-2.3.1-core.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-jdbc/2.3.1/hive-jdbc-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-llap-common/2.3.1/hive-llap-common-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-metastore/2.3.1/hive-metastore-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-serde/2.3.1/hive-serde-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-service/2.3.1/hive-service-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-service-rpc/2.3.1/hive-service-rpc-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/shims/hive-shims-0.23/2.3.1/hive-shims-0.23-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/shims/hive-shims-common/2.3.1/hive-shims-common-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-storage-api/2.3.1/hive-storage-api-2.3.1.jar wget https://repo1.maven.org/maven2/org/apache/hive/hive-shims/2.3.1/hive-shims-2.3.1.jar wget https://repo1.maven.org/maven2/org/json/json/20090211/json-20090211.jar cp /opt/hive/lib/log* /opt/spark/jars/ rm log4j-slf4j-impl-2.6.2.jar cd /opt
Copy the integration tests jar into the docker container
docker cp packaging/hudi-integ-test-bundle/target/hudi-integ-test-bundle-0.6.1-SNAPSHOT.jar adhoc-2:/opt
docker exec -it adhoc-2 /bin/bash
Clean the working directories before starting a new test:
hdfs dfs -rm -r /user/hive/warehouse/hudi-integ-test-suite/output/ hdfs dfs -rm -r /user/hive/warehouse/hudi-integ-test-suite/input/
Launch a Copy-on-Write job:
# COPY_ON_WRITE tables ========================= ## Run the following command to start the test suite spark-submit \ --packages org.apache.spark:spark-avro_2.11:2.4.0 \ --conf spark.task.cpus=1 \ --conf spark.executor.cores=1 \ --conf spark.task.maxFailures=100 \ --conf spark.memory.fraction=0.4 \ --conf spark.rdd.compress=true \ --conf spark.kryoserializer.buffer.max=2000m \ --conf spark.serializer=org.apache.spark.serializer.KryoSerializer \ --conf spark.memory.storageFraction=0.1 \ --conf spark.shuffle.service.enabled=true \ --conf spark.sql.hive.convertMetastoreParquet=false \ --conf spark.driver.maxResultSize=12g \ --conf spark.executor.heartbeatInterval=120s \ --conf spark.network.timeout=600s \ --conf spark.yarn.max.executor.failures=10 \ --conf spark.sql.catalogImplementation=hive \ --class org.apache.hudi.integ.testsuite.HoodieTestSuiteJob \ /opt/hudi-integ-test-bundle-0.6.1-SNAPSHOT.jar \ --source-ordering-field test_suite_source_ordering_field \ --use-deltastreamer \ --target-base-path /user/hive/warehouse/hudi-integ-test-suite/output \ --input-base-path /user/hive/warehouse/hudi-integ-test-suite/input \ --target-table table1 \ --props file:/var/hoodie/ws/docker/demo/config/test-suite/test.properties \ --schemaprovider-class org.apache.hudi.utilities.schema.FilebasedSchemaProvider \ --source-class org.apache.hudi.utilities.sources.AvroDFSSource \ --input-file-size 125829120 \ --workload-yaml-path file:/var/hoodie/ws/docker/demo/config/test-suite/complex-dag-cow.yaml \ --workload-generator-classname org.apache.hudi.integ.testsuite.dag.WorkflowDagGenerator \ --table-type COPY_ON_WRITE \ --compact-scheduling-minshare 1
Or a Merge-on-Read job:
# MERGE_ON_READ tables ========================= ## Run the following command to start the test suite spark-submit \ --packages org.apache.spark:spark-avro_2.11:2.4.0 \ --conf spark.task.cpus=1 \ --conf spark.executor.cores=1 \ --conf spark.task.maxFailures=100 \ --conf spark.memory.fraction=0.4 \ --conf spark.rdd.compress=true \ --conf spark.kryoserializer.buffer.max=2000m \ --conf spark.serializer=org.apache.spark.serializer.KryoSerializer \ --conf spark.memory.storageFraction=0.1 \ --conf spark.shuffle.service.enabled=true \ --conf spark.sql.hive.convertMetastoreParquet=false \ --conf spark.driver.maxResultSize=12g \ --conf spark.executor.heartbeatInterval=120s \ --conf spark.network.timeout=600s \ --conf spark.yarn.max.executor.failures=10 \ --conf spark.sql.catalogImplementation=hive \ --class org.apache.hudi.integ.testsuite.HoodieTestSuiteJob \ /opt/hudi-integ-test-bundle-0.6.1-SNAPSHOT.jar \ --source-ordering-field test_suite_source_ordering_field \ --use-deltastreamer \ --target-base-path /user/hive/warehouse/hudi-integ-test-suite/output \ --input-base-path /user/hive/warehouse/hudi-integ-test-suite/input \ --target-table table1 \ --props file:/var/hoodie/ws/docker/demo/config/test-suite/test.properties \ --schemaprovider-class org.apache.hudi.utilities.schema.FilebasedSchemaProvider \ --source-class org.apache.hudi.utilities.sources.AvroDFSSource \ --input-file-size 125829120 \ --workload-yaml-path file:/var/hoodie/ws/docker/demo/config/test-suite/complex-dag-mor.yaml \ --workload-generator-classname org.apache.hudi.integ.testsuite.dag.WorkflowDagGenerator \ --table-type MERGE_ON_READ \ --compact-scheduling-minshare 1
For long running test suite, validation has to be done differently. Idea is to run same dag in a repeated manner. Hence “ValidateDatasetNode” is introduced which will read entire input data and compare it with hudi contents both via spark datasource and hive table via spark sql engine.
If you have “ValidateDatasetNode” in your dag, do not replace hive jars as instructed above. Spark sql engine does not go well w/ hive2* jars. So, after running docker setup, just copy test.properties and your dag of interest and you are good to go ahead.
For repeated runs, two additional configs need to be set. “dag_rounds” and “dag_intermittent_delay_mins”. This means that your dag will be repeated for N times w/ a delay of Y mins between each round.
Also, ValidateDatasetNode can be configured in two ways. Either with “delete_input_data: true” set or not set. When “delete_input_data” is set for ValidateDatasetNode, once validation is complete, entire input data will be deleted. So, suggestion is to use this ValidateDatasetNode as the last node in the dag with “delete_input_data”. Example dag:
Insert
Upsert
ValidateDatasetNode with delete_input_data = true
If above dag is run with “dag_rounds” = 10 and “dag_intermittent_delay_mins” = 10, then this dag will run for 10 times with 10 mins delay between every run. At the end of every run, records written as part of this round will be validated. At the end of each validation, all contents of input are deleted.
For eg: incase of above dag,
Round1:
insert => inputPath/batch1
upsert -> inputPath/batch2
Validate with delete_input_data = true
Validates contents from batch1 and batch2 are in hudi and ensures Row equality
Since "delete_input_data" is set, deletes contents from batch1 and batch2.
Round2:
insert => inputPath/batch3
upsert -> inputPath/batch4
Validate with delete_input_data = true
Validates contents from batch3 and batch4 are in hudi and ensures Row equality
Since "delete_input_data" is set, deletes contents from batch3 and batch4.
Round3:
insert => inputPath/batch5
upsert -> inputPath/batch6
Validate with delete_input_data = true
Validates contents from batch5 and batch6 are in hudi and ensures Row equality
Since "delete_input_data" is set, deletes contents from batch5 and batch6.
.
.
If you wish to do a cumulative validation, do not set delete_input_data in ValidateDatasetNode. But remember that this may not scale beyond certain point since input data as well as hudi content's keeps occupying the disk and grows for every cycle.
Lets see an example where you don't set “delete_input_data” as part of Validation.
Round1:
insert => inputPath/batch1
upsert -> inputPath/batch2
Validate: validates contents from batch1 and batch2 are in hudi and ensures Row equality
Round2:
insert => inputPath/batch3
upsert -> inputPath/batch4
Validate: validates contents from batch1 to batch4 are in hudi and ensures Row equality
Round3:
insert => inputPath/batch5
upsert -> inputPath/batch6
Validate: validates contents from batch1 and batch6 are in hudi and ensures Row equality
.
.
You could also have validations in the middle of your dag and not set the “delete_input_data”. But set it only in the last node in the dag.
Round1:
insert => inputPath/batch1
upsert -> inputPath/batch2
Validate: validates contents from batch1 and batch2 are in hudi and ensures Row equality
insert => inputPath/batch3
upsert -> inputPath/batch4
Validate with delete_input_data = true
Validates contents from batch1 to batch4 are in hudi and ensures Row equality
since "delete_input_data" is set to true, this node deletes contents from batch1 and batch4.
Round2:
insert => inputPath/batch5
upsert -> inputPath/batch6
Validate: validates contents from batch5 and batch6 are in hudi and ensures Row equality
insert => inputPath/batch7
upsert -> inputPath/batch8
Validate: validates contents from batch5 to batch8 are in hudi and ensures Row equality
since "delete_input_data" is set to true, this node deletes contents from batch5 to batch8.
Round3:
insert => inputPath/batch9
upsert -> inputPath/batch10
Validate: validates contents from batch9 and batch10 are in hudi and ensures Row equality
insert => inputPath/batch11
upsert -> inputPath/batch12
Validate with delete_input_data = true
Validates contents from batch9 to batch12 are in hudi and ensures Row equality
Set "delete_input_data" to true. so this node deletes contents from batch9 to batch12.
.
.
Above dag was just an example for illustration purposes. But you can make it complex as per your needs.
Insert
Upsert
Delete
Validate w/o deleting
Insert
Rollback
Validate w/o deleting
Upsert
Validate w/ deletion
With this dag, you can set the two additional configs “dag_rounds” and “dag_intermittent_delay_mins” and have a long running test suite.
dag_rounds: 1
dag_intermittent_delay_mins: 10
dag_content:
Insert
Upsert
Delete
Validate w/o deleting
Insert
Rollback
Validate w/o deleting
Upsert
Validate w/ deletion
Sample COW command with repeated runs.
spark-submit \ --packages org.apache.spark:spark-avro_2.11:2.4.0 \ --conf spark.task.cpus=1 \ --conf spark.executor.cores=1 \ --conf spark.task.maxFailures=100 \ --conf spark.memory.fraction=0.4 \ --conf spark.rdd.compress=true \ --conf spark.kryoserializer.buffer.max=2000m \ --conf spark.serializer=org.apache.spark.serializer.KryoSerializer \ --conf spark.memory.storageFraction=0.1 \ --conf spark.shuffle.service.enabled=true \ --conf spark.sql.hive.convertMetastoreParquet=false \ --conf spark.driver.maxResultSize=12g \ --conf spark.executor.heartbeatInterval=120s \ --conf spark.network.timeout=600s \ --conf spark.yarn.max.executor.failures=10 \ --conf spark.sql.catalogImplementation=hive \ --conf spark.driver.extraClassPath=/var/demo/jars/* \ --conf spark.executor.extraClassPath=/var/demo/jars/* \ --class org.apache.hudi.integ.testsuite.HoodieTestSuiteJob \ /opt/hudi-integ-test-bundle-0.6.1-SNAPSHOT.jar \ --source-ordering-field test_suite_source_ordering_field \ --use-deltastreamer \ --target-base-path /user/hive/warehouse/hudi-integ-test-suite/output \ --input-base-path /user/hive/warehouse/hudi-integ-test-suite/input \ --target-table table1 \ --props test.properties \ --schemaprovider-class org.apache.hudi.utilities.schema.FilebasedSchemaProvider \ --source-class org.apache.hudi.utilities.sources.AvroDFSSource \ --input-file-size 125829120 \ --workload-yaml-path file:/var/hoodie/ws/docker/demo/config/test-suite/complex-dag-cow.yaml \ --workload-generator-classname org.apache.hudi.integ.testsuite.dag.WorkflowDagGenerator \ --table-type COPY_ON_WRITE \ --compact-scheduling-minshare 1
A ready to use dag is available under docker/demo/config/test-suite/ that could give you an idea for long running dags. cow-per-round-mixed-validate.yaml
As of now, “ValidateDatasetNode” uses spark data source and hive tables for comparison. Hence COW and real time view in MOR can be tested.