Copyright 2018 ABSA Group Limited

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

Hyperdrive - An extensible streaming ingestion pipeline on top of Apache Spark

Build Status

master	develop

What is Hyperdrive?

Hyperdrive is a configurable and scalable ingestion platform that allows data movement and transformation from streaming sources with exactly-once fault-tolerance semantics by using Apache Spark Structured Streaming.

In Hyperdrive, each ingestion is defined by the three components reader, transformer and writer. This separation allows adapting to different streaming sources and sinks, while reusing transformations common across multiple ingestion pipelines.

Motivation

Similar to batch processing, data ingestion pipelines are needed to process streaming data sources. While solutions for data pipelines exist, exactly-once fault-tolerance in streaming processing is an intricate problem and cannot be solved with the same strategies that exist for batch processing.

This is the gap the Hyperdrive aims to fill, by leveraging the exactly-once guarantee of Spark's Structured Streaming and by providing a flexible data pipeline.

Architecture

The data ingestion pipeline of Hyperdrive consists of four components: readers, transformers, writers.

• Readers define how to connect to sources, e.g. how to connect to Kafka in a secure cluster by providing security directives, which topic and brokers to connect to.
• Transformers define transformations to be applied to the decoded DataFrame, e.g. dropping columns.
• Writers define where DataFrames should be sent after the transformations, e.g. into HDFS as Parquet files.

Built-in components

• KafkaStreamReader - reads from a Kafka topic.
• ConfluentAvroDecodingTransformer - decodes the payload as Confluent Avro (through ABRiS), retrieving the schema from the specified Schema Registry. This transformer is capable of seamlessly handling whatever schemas the payload messages are using.
• ConfluentAvroEncodingTransformer - encodes the payload as Confluent Avro (through ABRiS), updating the schema to the specified Schema Registry. This transformer is capable of seamlessly handling whatever schema the dataframe is using.
• ColumnSelectorStreamTransformer - selects all columns from the decoded DataFrame.
• AddDateVersionTransformerStreamWriter - adds columns for ingestion date and an auto-incremented version number, to be used for partitioning.
• ParquetStreamWriter - writes the DataFrame as Parquet, in append mode.
• KafkaStreamWriter - writes to a Kafka topic.

Custom components

Custom components can be implemented using the Component Archetype following the API defined in the package za.co.absa.hyperdrive.ingestor.api

• A custom component has to be a class which extends either of the abstract classes StreamReader, StreamTransformer or StreamWriter
• The class needs to have a companion object which implements the corresponding trait StreamReaderFactory, StreamTransformerFactory or StreamWriterFactory
• The implemented components have to be packaged to a jar file, which can then be added to the classpath of the driver. To use a component, it has to be configured as described under Usage

After that, the new component will be able to be seamlessly invoked from the driver.

Usage

Hyperdrive has to be executed with Spark. Due to Spark-Kafka integration issues, it will only work with Spark 2.3 and higher.

How to run

git clone [email protected]:AbsaOSS/hyperdrive.git
mvn clean package

Given a configuration file has already been created, hyperdrive can be executed as follows:

spark-submit --class za.co.absa.hyperdrive.driver.drivers.PropertiesIngestionDriver driver/target/driver*.jar config.properties

Alternatively, configuration properties can also be passed as command-line arguments

spark-submit --class za.co.absa.hyperdrive.driver.drivers.CommandLineIngestionDriver driver/target/driver*.jar \
  component.ingestor=spark \
  component.reader=za.co.absa.hyperdrive.ingestor.implementation.reader.kafka.KafkaStreamReader \
  # more properties ...

Configuration

The configuration file may be created from the template located at driver/src/resources/Ingestion.properties.template.

CommandLineIngestionDriverDockerTest may be consulted for a working pipeline configuration.

General settings

Pipeline settings

Property Name	Required	Description
component.ingestor	Yes	Defines the ingestion pipeline. Only spark is currently supported.
component.reader	Yes	Fully qualified name of reader component, e.g.za.co.absa.hyperdrive.ingestor.implementation.reader.kafka.KafkaStreamReader
component.transformer.id.{order}	No	An arbitrary but unique string, referenced in this documentation as {transformer-id}
component.transformer.class.{transformer-id}	No	Fully qualified name of transformer component, e.g. za.co.absa.hyperdrive.ingestor.implementation.transformer.column.selection.ColumnSelectorStreamTransformer
component.writer	Yes	Fully qualified name of writer component, e.g. za.co.absa.hyperdrive.ingestor.implementation.writer.parquet.ParquetStreamWriter

Multiple transformers can be configured in the pipeline, including multiple instances of the same transformer. For each transformer instance, component.transformer.id.{order} and component.transformer.class.{transformer-id} have to specified, where {order} and {transformer-id} need to be unique. In the above table, {order} must be an integer and may be negative. {transformer-id} is only used within the configuration to identify which configuration options belong to a certain transformer instance.

Spark settings

Property Name	Required	Description
ingestor.spark.app.name	Yes	User-defined name of the Spark application. See Spark property spark.app.name
ingestor.spark.await.termination.timeout	No	Timeout in milliseconds. Stops query when timeout is reached. This option is only valid with termination method awaitTermination

Settings for built-in components

KafkaStreamReader

Property Name	Required	Description
reader.kafka.topic	Yes	The name of the kafka topic to ingest data from. Equivalent to Spark property subscribe
reader.kafka.brokers	Yes	List of kafka broker URLs . Equivalent to Spark property kafka.bootstrap.servers

Any additional properties for kafka can be added with the prefix reader.option.. E.g. the property kafka.security.protocol can be added as reader.option.kafka.security.protocol

See e.g. the Structured Streaming + Kafka Integration Guide for optional kafka properties.

ConfluentAvroStreamDecodingTransformer

The ConfluentAvroStreamDecodingTransformer is built on ABRiS. More details about the configuration properties can be found there. Caution: The ConfluentAvroStreamDecodingTransformer requires the property reader.kafka.topic to be set.

Property Name	Required	Description
transformer.{transformer-id}.schema.registry.url	Yes	URL of Schema Registry, e.g. http://localhost:8081. Equivalent to ABRiS property SchemaManager.PARAM_SCHEMA_REGISTRY_URL
transformer.{transformer-id}.value.schema.id	Yes	The schema id. Use latest or explicitly provide a number. Equivalent to ABRiS property SchemaManager.PARAM_VALUE_SCHEMA_ID
transformer.{transformer-id}.value.schema.naming.strategy	Yes	Subject name strategy of Schema Registry. Possible values are topic.name, record.name or topic.record.name. Equivalent to ABRiS property SchemaManager.PARAM_VALUE_SCHEMA_NAMING_STRATEGY
transformer.{transformer-id}.value.schema.record.name	Yes for naming strategies record.name and topic.record.name	Name of the record. Equivalent to ABRiS property SchemaManager.PARAM_SCHEMA_NAME_FOR_RECORD_STRATEGY
transformer.{transformer-id}.value.schema.record.namespace	Yes for naming strategies record.name and topic.record.name	Namespace of the record. Equivalent to ABRiS property SchemaManager.PARAM_SCHEMA_NAMESPACE_FOR_RECORD_STRATEGY
transformer.{transformer-id}.consume.keys	No	If set to true, keys will be consumed and added as columns to the dataframe. Key columns will be prefixed with key__
transformer.{transformer-id}.key.schema.id	Yes if consume.keys is true	The schema id for the key.
transformer.{transformer-id}.key.schema.naming.strategy	Yes if consume.keys is true	Subject name strategy for key
transformer.{transformer-id}.key.schema.record.name	Yes for key naming strategies record.name and topic.record.name	Name of the record.
transformer.{transformer-id}.key.schema.record.namespace	Yes for key naming strategies record.name and topic.record.name	Namespace of the record.

For detailed information on the subject name strategy, please take a look at the Schema Registry Documentation.

ConfluentAvroStreamEncodingTransformer

The ConfluentAvroStreamEncodingTransformer is built on ABRiS. More details about the configuration properties can be found there. Caution: The ConfluentAvroStreamEncodingTransformer requires the property writer.kafka.topic to be set.

Property Name	Required	Description
transformer.{transformer-id}.schema.registry.url	Yes	URL of Schema Registry, e.g. http://localhost:8081. Equivalent to ABRiS property SchemaManager.PARAM_SCHEMA_REGISTRY_URL
transformer.{transformer-id}.value.schema.naming.strategy	Yes	Subject name strategy of Schema Registry. Possible values are topic.name, record.name or topic.record.name. Equivalent to ABRiS property SchemaManager.PARAM_VALUE_SCHEMA_NAMING_STRATEGY
transformer.{transformer-id}.value.schema.record.name	Yes for naming strategies record.name and topic.record.name	Name of the record. Equivalent to ABRiS property SchemaManager.PARAM_SCHEMA_NAME_FOR_RECORD_STRATEGY
transformer.{transformer-id}.value.schema.record.namespace	Yes for naming strategies record.name and topic.record.name	Namespace of the record. Equivalent to ABRiS property SchemaManager.PARAM_SCHEMA_NAMESPACE_FOR_RECORD_STRATEGY
transformer.{transformer-id}.produce.keys	No	If set to true, keys will be produced according to the properties key.column.prefix and key.column.names of the Hyperdrive Context
transformer.{transformer-id}.key.schema.naming.strategy	Yes if produce.keys is true	Subject name strategy for key
transformer.{transformer-id}.key.schema.record.name	Yes for key naming strategies record.name and topic.record.name	Name of the record.
transformer.{transformer-id}.key.schema.record.namespace	Yes for key naming strategies record.name and topic.record.name	Namespace of the record.

ColumnSelectorStreamTransformer

Property Name	Required	Description
transformer.{transformer-id}.columns.to.select	Yes	Comma-separated list of columns to select. * can be used to select all columns. Only existing columns using column names may be selected (i.e. expressions cannot be constructed)

AddDateVersionTransformer

The AddDateVersionTransformer adds the columns hyperdrive_date and hyperdrive_version. hyperdrive_date is the ingestion date (or a user-defined date), while hyperdrive_version is a number automatically incremented with every ingestion, starting at 1. For the auto-increment to work, hyperdrive_date and hyperdrive_version need to be defined as partition columns. Caution: This transformer requires a writer which defines writer.parquet.destination.directory.

Property Name	Required	Description
transformer.{transformer-id}.report.date	No	User-defined date for hyperdrive_date in format yyyy-MM-dd. Default date is the date of the ingestion

See Pipeline settings for details about {transformer-id}.

ParquetStreamWriter

Property Name	Required	Description
writer.parquet.destination.directory	Yes	Destination path of the sink. Equivalent to Spark property path for the DataStreamWriter
writer.parquet.partition.columns	No	Comma-separated list of columns to partition by.
writer.parquet.metadata.check	No	Set this option to true if the consistency of the metadata log should be checked prior to the query. For very large tables, the check may be very expensive
writer.common.trigger.type	No	See Combination writer properties
writer.common.trigger.processing.time	No	See Combination writer properties

Any additional properties for the DataStreamWriter can be added with the prefix writer.parquet.options, e.g. writer.parquet.options.key=value

KafkaStreamWriter

Property Name	Required	Description
writer.kafka.topic	Yes	The name of the kafka topic to ingest data from. Equivalent to Spark property topic
writer.kafka.brokers	Yes	List of kafka broker URLs . Equivalent to Spark property kafka.bootstrap.servers
writer.common.trigger.type	No	See Combination writer properties
writer.common.trigger.processing.time	No	See Combination writer properties

Common writer properties

Property Name	Required	Description
writer.common.checkpoint.location	Yes	Used for Spark property checkpointLocation. The checkpoint location has to be unique among different workflows.
writer.common.trigger.type	No	Either Once for one-time execution or ProcessingTime for micro-batch executions for micro-batch execution. Default: Once.
writer.common.trigger.processing.time	No	Interval in ms for micro-batch execution (using ProcessingTime). Default: 0ms, i.e. execution as fast as possible.

Behavior of Triggers

Trigger (writer.common.trigger.type)	Timeout (ingestor.spark.termination.timeout)	Runtime	Details
Once	No timeout	Limited	Consumes all data that is available at the beginning of the micro-batch. The query processes exactly one micro-batch and stops then, even if more data would be available at the end of the micro-batch.
ProcessingTime	With timeout	Limited	Consumes data in micro-batches and only stops when the timeout is reached or the query is killed.
ProcessingTime	No timeout	Long-running	Consumes data in micro-batches and only stops when the query is killed.

• Note 1: The first micro-batch of the query will contain all available messages to consume and can therefore be quite large, even if the trigger ProcessingTime is configured, and regardless of what micro-batch interval is configured. To limit the size of a micro-batch, the property reader.option.maxOffsetsPerTrigger should be used. See also http://spark.apache.org/docs/latest/structured-streaming-kafka-integration.html
• Note 2: It's possible to define a timeout for trigger Once. If the timeout is reached before the micro-batch is processed, it won't be completed and no data will be committed. Such a behavior seems quite unpredictable and therefore we don't recommend it.

See the Spark Documentation for more information about triggers.

Hyperdrive Context

HyperdriveContext is an object intended to be used by the components to share data. It is a key-value store, where the key is a string and the value can be of any type. The following context variables are currently used by the default implementation.

Name	Type	Description
key.column.prefix	String	If ConfluentAvroDecodingTransformer is configured to consume keys, it prefixes the key columns with key__ such that they can be distinguished in the dataframe. If key__ happens to be a prefix of a value column, a random alphanumeric string is used instead.
key.column.names	Seq[String]	If ConfluentAvroDecodingTransformer is configured to consume keys, it contains the original column names (without prefix) in the key schema.

Other

Hyperdrive uses Apache Commons Configuration 2. This allows properties to be referenced, e.g. like so

transformer.[avro.decoder].schema.registry.url=http://localhost:8081
writer.kafka.schema.registry.url=${transformer.[avro.decoder].schema.registry.url}

Workflow Manager

Hyperdrive ingestions may be triggered using the Workflow Manager, which is developed in a separate repository: AbsaOSS/hyperdrive-trigger

A key feature of the Workflow Manager are triggers, which define when an ingestion should be executed and how it should be requested. The workflow manager supports cron-based triggers as well as triggers that listen to a notification topic.