yotpoltd / metorikku

A simplified, lightweight ETL Framework based on Apache Spark

GitHub

Metorikku Logo

Build Status

Metorikku is a library that simplifies writing and executing ETLs on top of Apache Spark.

It is based on simple YAML configuration files and runs on any Spark cluster.

The platform also includes a simple way to write unit and E2E tests.

Getting started

To run Metorikku you must first define 2 files.

Metric file

A metric file defines the steps and queries of the ETL as well as where and what to output.

For example a simple configuration YAML (JSON is also supported) should be as follows:

steps:
- dataFrameName: df1
  sql: 
    SELECT *
    FROM input_1
    WHERE id > 100
- dataFrameName: df2
  sql: 
    SELECT *
    FROM df1
    WHERE id < 1000
output:
- dataFrameName: df2
  outputType: Parquet
  outputOptions:
    saveMode: Overwrite
    path: df2.parquet

You can check out a full example file for all possible values in the sample YAML configuration file.

Make sure to also check out the full Spark SQL Language manual for the possible queries.

Job file

This file will include input sources, output destinations and the location of the metric config files.

So for example a simple YAML (JSON is also supported) should be as follows:

metrics:
  - /full/path/to/your/metric/file.yaml
inputs:
  input_1: parquet/input_1.parquet
  input_2: parquet/input_2.parquet
output:
    file:
        dir: /path/to/parquet/output

You can check out a full example file for all possible values in the sample YAML configuration file.

Also make sure to check out all our examples.

Supported input/output:

Currently Metorikku supports the following inputs: CSV, JSON, parquet, JDBC, Kafka, Cassandra

And the following outputs: CSV, JSON, parquet, Redshift, Cassandra, Segment, JDBC, Kafka, Elasticsearch

Running Metorikku

There are currently 3 options to run Metorikku.

Run on a spark cluster

To run on a cluster Metorikku requires Apache Spark v2.2+

  • Download the last released JAR
  • Run the following command: spark-submit --class com.yotpo.metorikku.Metorikku metorikku.jar -c config.yaml

Run locally

Metorikku is released with a JAR that includes a bundled spark.

  • Download the last released Standalone JAR
  • Run the following command: java -Dspark.master=local[*] -cp metorikku-standalone.jar com.yotpo.metorikku.Metorikku -c config.yaml
  • Also job in a JSON format is supported, run following command: java -Dspark.master=local[*] -cp metorikku-standalone.jar com.yotpo.metorikku.Metorikku --job "{*}"

Run as a library

It's also possible to use Metorikku inside your own software

Metorikku library requires scala 2.11

To use it add the following dependency to your build.sbt: "com.yotpo" % "metorikku" % "LATEST VERSION"

Metorikku Tester

In order to test and fully automate the deployment of metrics we added a method to run tests against a metric.

A test is comprised of the following:

Test settings

This defines what to test and where to get the mocked data. For example, a simple test YAML (JSON is also supported) will be:

metric: "/path/to/metric"
mocks:
- name: table_1
  path: mocks/table_1.jsonl
tests:
  df2:
  - id: 200
    name: test
  - id: 300
    name: test2

And the corresponding mocks/table_1.jsonl:

{ "id": 200, "name": "test" }
{ "id": 300, "name": "test2" }
{ "id": 1, "name": "test3" }

Running Metorikku Tester

You can run Metorikku tester in any of the above methods (just like a normal Metorikku).

The main class changes from com.yotpo.metorikku.Metorikku to com.yotpo.metorikku.MetorikkuTester

Testing streaming metrics

In Spark some behaviors are different when writing queries for streaming sources (for example kafka).

In order to make sure the test behaves the same as the real life queries, you can configure a mock to behave like a streaming input by writing the following:

metric: "/path/to/metric"
mocks:
- name: table_1
  path: mocks/table_1.jsonl
  # default is false
  streaming: true
# default is append output mode
outputMode: update
tests:
  df2:
  - id: 200
    name: test
  - id: 300
    name: test2

Notes

Variable interpolation

All configuration files support variable interpolation from environment variables and system properties using the following format: ${variable_name}

Using JDBC

When using JDBC writer or input you must provide a path to the driver JAR.

For example to run with spark-submit with a mysql driver: spark-submit --driver-class-path mysql-connector-java-5.1.45.jar --jars mysql-connector-java-5.1.45.jar --class com.yotpo.metorikku.Metorikku metorikku.jar -c config.yaml

If you want to run this with the standalone JAR: java -Dspark.master=local[*] -cp metorikku-standalone.jar:mysql-connector-java-5.1.45.jar -c config.yaml

JDBC query

JDBC query output allows running a query for each record in the dataframe.

Mandatory parameters:
  • query - defines the SQL query. In the query you can address the column of the DataFrame by their location using the dollar sign ($) followed by the column index. For example:
INSERT INTO table_name (column1, column2, column3, ...) VALUES ($1, $2, $3, ...);
Optional Parameters:
  • maxBatchSize - The maximum size of queries to execute against the DB in one commit.
  • minPartitions - Minimum partitions in the DataFrame - may cause repartition.
  • maxPartitions - Maximum partitions in the DataFrame - may cause coalesce.

Kafka output

Kafka output allows writing batch operations to kafka

We use spark-sql-kafka-0-10 as a provided jar - spark-submit command should look like so:

spark-submit --packages org.apache.spark:spark-sql-kafka-0-10_2.11:2.2.0 --class com.yotpo.metorikku.Metorikku metorikku.jar

Mandatory parameters:
  • topic - defines the topic in kafka which the data will be written to. currently supported only one topic

  • valueColumn - defines the values which will be written to the Kafka topic, Usually a json version of data, For example:

SELECT keyColumn, to_json(struct(*)) AS valueColumn FROM table
Optional Parameters:
  • keyColumn - key that can be used to perform de-duplication when reading

Streaming Input

Using streaming input will convert your application into a streaming application build on top of Spark Structured Streaming.

To enable all other writers and also enable multiple outputs for a single streaming dataframe, add batchMode to your job configuration, this will enable the foreachBatch mode (only available in spark >= 2.4.0) Check out all possible streaming configurations in the streaming section of the sample job configuration file.

Please note the following while using streaming applications:

  • Multiple streaming aggregations (i.e. a chain of aggregations on a streaming DF) are not yet supported on streaming Datasets.

  • Limit and take first N rows are not supported on streaming Datasets.

  • Distinct operations on streaming Datasets are not supported.

  • Sorting operations are supported on streaming Datasets only after an aggregation and in Complete Output Mode.

  • Make sure to add the relevant Output Mode to your Metric as seen in the Examples

  • Make sure to add the relevant Triggers to your Metric if needed as seen in the Examples

  • For more information please go to Spark Structured Streaming WIKI

Kafka Input

Kafka input allows reading messages from topics

inputs:
  testStream:
    kafka:
      servers:
        - 127.0.0.1:9092
      topic: test
      consumerGroup: testConsumerGroupID # optional
      schemaRegistryUrl: https://schema-registry-url # optional
      schemaSubject: subject # optional

When using kafka input, writing is only available to File and Kafka, and only to a single output.

  • In order to measure your consumer lag you can use the consumerGroup parameter to track your application offsets against your kafka input. This will commit the offsets to kafka, as a new dummy consumer group.

  • In order to deserialize your kafka stream messages with confluent's Schema Registry, add the schemaRegistryUrl option to the kafka input config

  • If your subject schema name is not <TOPIC NAME>-value (e.g. if the topic is a regex pattern) you can specify the schema subject in the schemaSubject section

Topic Pattern

Kafka input also allows reading messages from multiple topics by using subscribe pattern:

inputs:
  testStream:
    kafka:
      servers:
        - 127.0.0.1:9092
      # topicPattern can be any Java regex string
      topicPattern: my_topics_regex.*
      consumerGroup: testConsumerGroupID # optional
      schemaRegistryUrl: https://schema-registry-url # optional
      schemaSubject: subject # optional
  • While using topicPattern, consider using schemaRegistryUrl and schemaSubject in case your topics have different schemas.
File Streaming Input

Metorikku supports streaming over a file system as well. You can use the Data stream reading by specifying isStream: true, and a specific path in the job (must be a single path) as a streaming source, this will trigger jobs for new files added to the folder.

inputs:
  testStream:
    file:
      path: examples/file_input_stream/input
      isStream: true
      format: json
      options:
        timestampFormat: "yyyy-MM-dd'T'HH:mm:ss'Z'"
Watermark

Metorikku supports Watermark method which helps a stream processing engine to deal with late data. You can use watermarking by adding a new udf step in your metric:

# This will become the new watermarked dataframe name.
- dataFrameName: dataframe
  classpath: com.yotpo.metorikku.code.steps.Watermark
  params:
    # Watermark table my_table
    table: my_table
    # The column representing the event time (needs to be a TIMESTAMP or DATE column)
    eventTime: event
    delayThreshold: 2 hours

Instrumentation

One of the most useful features in Metorikku is it's instrumentation capabilities.

Instrumentation metrics are written by default to what's configured in spark-metrics.

Metorikku sends automatically on top of what spark is already sending the following:

  • Number of rows written to each output

  • Number of successful steps per metric

  • Number of failed steps per metric

  • In streaming: records per second

  • In streaming: number of processed records in batch

You can also send any information you like to the instrumentation output within a metric. by default the last column of the schema will be the field value. Other columns that are not value or time columns will be merged together as the name of the metric. If writing directly to influxDB these will become tags.

Check out the example for further details.

using InfluxDB

You can also send metric directly to InfluxDB (gaining the ability to use tags and time field).

Check out the example and also the InfluxDB E2E test for further details.

Elasticsearch output

Elasticsearch output allows bulk writing to elasticsearch We use elasticsearch-hadoop as a provided jar - spark-submit command should look like so:

spark-submit --packages org.elasticsearch:elasticsearch-hadoop:6.6.1 --class com.yotpo.metorikku.Metorikku metorikku.jar

Check out the example and also the Elasticsearch E2E test for further details.

Docker

Metorikku is provided with a docker image.

You can use this docker to deploy metorikku in container based environments (we're using Nomad by HashiCorp).

Check out this docker-compose for a full example of all the different parameters available and how to set up a cluster.

Currently the image only supports running metorikku in a spark cluster mode with the standalone scheduler.

The image can also be used to run E2E tests of a metorikku job. Check out an example of running a kafka 2 kafka E2E with docker-compose here

UDF

Metorikku supports adding custom code as a step. This requires creating a JAR with the custom code. Check out the UDF examples directory for a very simple example of such a JAR.

The only thing important in this JAR is that you have an object with the following method:

object SomeObject {
  def run(ss: org.apache.spark.sql.SparkSession, metricName: String, dataFrameName: String, params: Option[Map[String, String]]): Unit = {}
}

Inside the run function do whatever you feel like, in the example folder you'll see that we registered a new UDF. Once you have a proper scala file and a build.sbt file you can run sbt package to create the JAR.

When you have the newly created JAR (should be in the target folder), copy it to the spark cluster (you can of course also deploy it to your favorite repo).

You must now include this JAR in your spark-submit command by using the --jars flag, or if you're using java to run add it to the -cp flag.

Now all that's left is to add it as a new step in your metric:

- dataFrameName: dataframe
  classpath: com.example.SomeObject
  params:
    param1: value1

This will trigger your run method with the above dataFrameName. Check out the built-in code steps here.

NOTE: If you added some dependencies to your custom JAR build.sbt you have to either use sbt-assembly to add them to the JAR or you can use the --packages when running the spark-submit command

Apache Hive metastore

Metorikku supports reading and saving tables with Apache hive metastore. To enable hive support via spark-submit (assuming you're using MySQL as Hive's DB but any backend can work) send the following configurations:

spark-submit \
--packages mysql:mysql-connector-java:5.1.75 \
--conf spark.sql.catalogImplementation=hive \
--conf spark.hadoop.javax.jdo.option.ConnectionURL="jdbc:mysql://localhost:3306/hive?useSSL=false&createDatabaseIfNotExist=true" \
--conf spark.hadoop.javax.jdo.option.ConnectionDriverName=com.mysql.jdbc.Driver \
--conf spark.hadoop.javax.jdo.option.ConnectionUserName=user \
--conf spark.hadoop.javax.jdo.option.ConnectionPassword=pass \
--conf spark.sql.warehouse.dir=/warehouse ...

NOTE: If you're running via the standalone metorikku you can use system properties instead (-Dspark.hadoop...) and you must add the MySQL connector JAR to your class path via -cp

This will enable reading from the metastore.

To write an external table to the metastore you need to add tableName to your output configuration:

...
output:
- dataFrameName: moviesWithRatings
  outputType: Parquet
  outputOptions:
    saveMode: Overwrite
    path: moviesWithRatings.parquet
    tableName: hiveTable
    overwrite: true

Only file formats are supported for table saves (Parquet, CSV, JSON).

To write a managed table (that will reside in the warehouse dir) simply omit the path in the output configuration.

To change the default database you can add the following to the job configuration:

...
catalog:
  database: some_database
...

Check out the examples and the E2E test

Apache Hudi

Metorikku supports reading/writing with Apache Hudi.

Hudi is a very exciting project that basically allows upserts and deletes directly on top of partitioned parquet data.

In order to use Hudi with Metorikku you need to add to your classpath (via --jars or if running locally with -cp) an external JAR from here: http://central.maven.org/maven2/com/uber/hoodie/hoodie-spark-bundle/0.4.7/hoodie-spark-bundle-0.4.7.jar

To run Hudi jobs you also have to make sure you have the following spark configuration (pass with --conf or -D):

spark.serializer=org.apache.spark.serializer.KryoSerializer

After that you can start using the new Hudi writer like this:

Job config

output:
  hudi:
    dir: /examples/output
    # This controls the level of parallelism of hudi writing (should be similar to shuffle partitions)
    parallelism: 1
    # upsert/insert/bulkinsert
    operation: upsert
    # COPY_ON_WRITE/MERGE_ON_READ
    storageType: COPY_ON_WRITE
    # Maximum number of versions to retain 
    maxVersions: 1
    # Hive database to use when writing
    hiveDB: default
    # Hive server URL
    hiveJDBCURL: jdbc:hive2://hive:10000
    hiveUserName: root
    hivePassword: pass

Metric config

dataFrameName: test
  outputType: Hudi
  outputOptions:
    path: test.parquet
    # The key to use for upserts
    keyColumn: userkey
    # This will be used to determine which row should prevail (newer timestamps will win)
    timeColumn: ts
    # Partition column - note that hudi support a single column only, so if you require multiple levels of partitioning you need to add / to your column values
    partitionBy: date
    # Mapping of the above partitions to hive (for example if above is yyyy/MM/dd than the mapping should be year,month,day)
    hivePartitions: year,month,day
    # Hive table to save the results to
    tableName: test_table

In order to delete send in your dataframe a boolean column called _hoodie_delete, if it's true that row will be deleted.

Check out the examples and the E2E test for more details.

Also check the full list of configurations possible with hudi here.

License

See the LICENSE file for license rights and limitations (MIT).