ApacheCon North America 2018: Creating Spark Data Sources
Download as PPTX, PDF1 like267 views
This document discusses building custom Spark data sources and provides an overview of the Spark data source API version 2. Building a custom data source allows exploiting special features of data sources that are not supported by existing built-in or third party data sources. The document outlines reasons for building a custom data source, the key interfaces that must be implemented as part of the data source API, and a code walkthrough of a simple example data source. It also discusses practical considerations like understanding the characteristics of the target data source. The data source API is still evolving so custom data sources may need to be updated for future Spark versions.
ApacheCon North America 2018: Creating Spark Data Sources
- 1. Under-the-hood: Creating Your Own Spark Data Sources Speaker: Jayesh Thakrar @ Conversant Slide 1
- 2. Why Build Your Own Data Source? Slide 2 Don't need • text, csv, ORC, Parquet, JSON files • JDBC sources • When available from project / vendor, e.g. • Cassandra, Kafka, MongoDB, etc • Teradata, Greenplum (2018), etc Need to • Use special features, e.g. • Kafka transactions • Exploit RDBMS specific partition features • Because you can, and want to Conversant Use Cases (mid-2017) • Greenplum as data source (read/write) • Incompatibility between Spark, Kafka and Kafka connector versions
- 3. Agenda 1. Introduction to Spark data sources 2. Walk-through sample code 3. Practical considerations Slide 3
- 4. Introduction To Spark Data Source Slide 4
- 5. Using Data Sources Slide 5 • Built-in spark.read.csv("path") spark.read.orc("path") spark.read.parquet("path") • Third-party/custom spark.read.format("class-name").load() // custom data source spark.read.format("...").option("...", "...").load() // with options spark.read.format("...").schema("..").load() // with schema
- 6. Spark Data Source Spark Application Spark API Spark Data SourceData Spark Backend (runtime, Spark libraries, etc) Slide 6
- 7. Data Source V2 API Spark Application Spark API Spark Data SourceData Spark Backend (runtime, Spark libraries, etc) Shiny, New V2 API since Spark 2.3 (Feb 2018) SPARK-15689 No relations, scans involved Slide 7
- 8. V2 API: Data Source Types Data Source Type Output Batch Dataset Microbatch (successor to Dstreams) Structured Stream = stream of bounded dataset Continuous Continuous Stream = continuous stream of Row(s) Slide 8
- 9. V2 DATA SOURCE API • Well documented design V2 API : Design Doc , Feature SPARK-15689 Continuous Streaming Design Doc and Feature - SPARK-20928 • Implemented as Java Interfaces (not classes) • Similar interfaces across all data source types • Microbatch and Continuous not hardened yet... e.g. SPARK-23886, SPARK-23887, SPARK-22911 Slide 9
- 11. Details • Project: https://github.com/JThakrar/sparkconn • Requirements: • Scala 2.11 • SBT Slide 11
- 12. Reading From Data Source val data = spark.read.format("DataSource").option("key", "value").schema("..").load() Step Action spark = SparkSession read Returns a DataFrameReader for data source orchestration format Lazy lookup of data source class - by shortname or by full-qualified class name option Zero, one or more key-value pairs of options for data source schema Optional, user-provided schema load Loads the data as a dataframe. Remember dataframes are lazily-evaluated. Slide 12
- 13. Reading: Interfaces to Implement Slide 13 spark.format(...) DataSourceRegister ReadSupport ReadSupportWithSchema DataSourceReader DataReaderFactory DataReaderFactory DataReaderFactory DataReader DataReader DataReader Driver Executors / Tasks / Partitions Need to implement a minimum of 5 interfaces and 3 classes
- 14. Interface Definitions And Dependency Slide 14 libraryDependencies += "org.apache.spark" %% "spark-sql" % "2.3.1" % "provided"
- 15. Example V2 API Based Datasource Slide 15 • Very simple data source that generates a row with a fixed schema of a single string column. Column name = "string_value" • Completely self-contained (i.e. no external connection) • Number of partitions in dataset is user-configurable (default = 5) • All partitions contains same number of rows (strings) • Number of rows per partition is user-configurable (default = 5)
- 16. Read Interface: DataSourceRegister Slide 16 Interface Purpose org.apache.spark.sql.sources.v2.DataSourceRegister org.apache.spark.sql.sources.v2.ReadSupport and / or org.apache.spark.sql.sources.v2.ReadSupportWithSchema DataSourceRegister is the entry point for your data source. ReadSupport is then responsible to instantiate the object implementing DataSourceReader (discussed later). It accepts the options/parameters and optional schema from Spark application.
- 17. Read Interface: DataSourceReader Slide 17 Interface Purpose org.apache.spark.sql.sources.v2.reader.DataSourceReader This interface requires implementations for: • determining the schema of data • determining the number of partitions and creating that many reader factories below.
- 18. Read Interface: DataReaderFactory Slide 18 Interface Purpose org.apache.spark.sql.sources.v2.reader.DataReaderFactory This is the "handle" passed by the driver to each executor. It instantiates the readers below and controls data fetch.
- 19. Read Interface: DataReader Slide 19 Interface Purpose org.apache.spark.sql.sources.v2.reader.DataReader This does the actual work of fetching data from source (at task-level)
- 20. Summary Slide 20 Interface Purpose org.apache.spark.sql.sources.v2.DataSourceRegister org.apache.spark.sql.sources.v2.ReadSupport org.apache.spark.sql.sources.v2.ReadSupportWithSchema DataSourceRegister is the entry point for your connector. ReadSupport is then responsible to instantiate the object implementing DataSourceReader below. It accepts the options/parameters and optional schema from Spark application. org.apache.spark.sql.sources.v2.reader.DataSourceReader This interface requires implementations for: • determining the schema of data • determining the number of partitions and creating that many reader factories below. The DataSourceRegister, DataSourceReader and DataReaderFactory are instantiated at the driver. The driver then serializes DataReaderFactory and sends it to each of the executors. org.apache.spark.sql.sources.v2.reader.DataReaderFactory This is the "handle" passed by the driver to each executor. It instantiates the readers below and controls data fetch. org.apache.spark.sql.sources.v2.reader.DataReader This does the actual work of fetching data Because you are implementing interfaces, YOU can determine the "class" parameters and initialization
- 21. Write Interfaces to Implement Slide 21 Interface Purpose org.apache.spark.sql.sources.v2.DataSourceRegister org.apache.spark.sql.sources.v2.WriteSupport DataSourceRegister is the entry point for your connector. WriteSupport is then responsible to instantiate the object implementing DataSourceWriter below. It accepts the options/parameters and the schema. org.apache.spark.sql.sources.v2.writer.DataSourceWriter This interface requires implementations for: • committing data write • aborting data write • creating writer factories The DataSourceRegister, DataSourceWriter and DataWriterFactory are instantiated at the driver. The driver then serializes DataWriterFactory and sends it to each of the executors. org.apache.spark.sql.sources.v2.writer.DataWriterFactory This is the "handle" passed by the driver to each executor. It instantiates the writers below and controls data fetch. org.apache.spark.sql.sources.v2.writer.DataWriter This does the actual work of writing and committing/aborting the data org.apache.spark.sql.sources.v2.writer.WriterCommitMessage This is a "commit" message that is passed from the DataWriter to DataSourceWriter.
- 23. Know Your Data Source • Configuration • Partitions • Data schema • Parallelism approach • Batch and/or streaming • Restart / recovery Slide 23
- 24. V2 API IS STILL EVOLVING • SPARK-22386 - Data Source V2 Improvements • SPARK-23507 - Migrate existing data sources • SPARK-24073 - DataReaderFactory Renamed in 2.4 • SPARK-24252, SPARK-25006 - DataSourceV2: Add catalog support • So Why use V2? Future-ready and alternative to V2 needs significantly more time and effort! See https://www.youtube.com/watch?v=O9kpduk5D48 Slide 24
- 25. About... Conversant • Digital marketing unit of Epsilon under Alliance Data Systems (ADS) • (Significant) player in internet advertising. We see about 80% of internet ad bids in the US • Secret sauce = anonymous cross-device profiles driving personalized messaging Me (Jayesh Thakrar) • Sr. Software Engineer (jthakrar@conversantmedia.com) • https://www.linkedin.com/in/jayeshthakrar/ Slide 25