Extending Apache Spark APIs Without Going Near Spark Source or a Compiler with Anna Holschuh
0 likes639 views
The document discusses how to extend Apache Spark APIs without modifying Spark source code using Scala's "Enrich My Library" pattern. It provides an example of adding a .validate() method to Dataset objects to enable validation checks. The pattern involves defining an implicit class that augments existing types with new methods. This allows validation classes to integrate seamlessly with Spark jobs while keeping code concise, isolated and testable. Other uses like metrics collection and logging are also discussed.
![An Example
18#DevSAIS19
Step 3: Define a validation
• Our validation extends a Validation typed with
Dataset[Article]
• It fills in the abstract method .execute() which
defines what the validation is checking for
• This means that any time the compiler finds a
Dataset[Article] type, we can call .validate() on
it with this validation supplied because of our
implicit class
• Roughly 20 lines of concise and isolated code
is nicely separated from the core ETL job](https://image.slidesharecdn.com/9annaholschuh-180615013932/85/Extending-Apache-Spark-APIs-Without-Going-Near-Spark-Source-or-a-Compiler-with-Anna-Holschuh-18-320.jpg)
Extending Apache Spark APIs Without Going Near Spark Source or a Compiler with Anna Holschuh
- 1. Anna Holschuh, Target Extending Apache Spark APIs Without Going Near Spark Source or a Compiler #DevSAIS19
- 2. What This Talk is About • Scala programming constructs • Functional programming paradigms • Tips for organizing code in production systems 2#DevSAIS19
- 3. Who am I • Lead Data Engineer at Target since 2016 • Deep love of all things Target • Primary career focus has been building backend systems with a personal passion for Machine Learning problems • Started working in Spark in 2015 3#DevSAIS19
- 4. Agenda • Motivation • Scala’s “Enrich My Library” Pattern • An Example • Other Uses 4#DevSAIS19
- 5. Agenda • Motivation • Scala’s “Enrich My Library” Pattern • An Example • Other Uses 5#DevSAIS19
- 6. Motivation Let’s go through an example… • We have a system of Authors, Articles, and Comments on those Articles • From the example, Spark/Scala lends itself well to functional programming paradigms • What happens when the system grows in size/complexity and it becomes necessary to inject more custom code into the mix? • Can we keep things concise, readable, and efficient using the same functional style of code development? 6#DevSAIS19
- 7. Motivation Functional Programming Refresher • Declarative style of writing code (vs. Imperative) • Favors composition with functions • Avoids shared state, mutability, and side effects. 7#DevSAIS19
- 8. Motivation A Validation Framework was born… • Tasked with building an on-demand computation system consuming various data sources • There were many ways for this data to go wrong • Needed a way to fail fast and in a predictable way when a certain bar for quality was not being met 8#DevSAIS19
- 9. Motivation A Validation Framework was born… • Desired ability to “sprinkle” .validate() calls throughout our existing Spark ETL code 9#DevSAIS19 This is possible with Scala’s “Enrich My Library” Pattern
- 10. Agenda • Motivation • Scala’s “Enrich My Library” Pattern • An Example • Other Uses 10#DevSAIS19
- 11. “Enrich My Library” A Scala programming pattern… • Allows us to augment existing APIs • Analogous features in other languages • Also known as “Pimp My Library” for Googling purposes • Syntactic sugar that uses implicit classes to guide the compiler 11#DevSAIS19 Reference: https://docs.scala-lang.org/overviews/core/implicit-classes.html
- 12. “Enrich My Library” What are implicits? Scala supports a keyword “implicit” that allows the compiler to implicitly make connections at compile-time as opposed to explicitly having to call a function or feed in a variable. Scala supports implicit values, parameters, functions, and classes. What is an implicit class? Introduced formally with Scala 2.10 although it’s possible to achieve the same effect in previous versions through different constructs. Allows extension of classes one normally wouldn’t have access to. 12#DevSAIS19 Reference: https://docs.scala-lang.org/overviews/core/implicit-classes.html
- 13. Agenda • Motivation • Scala’s “Enrich My Library” Pattern • An Example • Other Uses 13#DevSAIS19
- 14. An Example Back to our example… How do we go from THIS 14#DevSAIS19
- 15. Motivation 15#DevSAIS19 Back to our example… To THIS
- 16. An Example 16#DevSAIS19 Step 1: Build a Validation class to work with • Abstract class parameterized with type T representing the object type that we plan to validate • Contains metadata relevant to running a validation • Has an abstract .execute() method to be filled in by concrete subclasses • Contains a concrete implementation .performValidation() that calls on the abstract execute method
- 17. An Example 17#DevSAIS19 Step 2: Add an implicit class to allow the decoration of existing types with new methods • The class can be named anything • It must be nested in a package or object • It can only have one parameter. This defines what class it’s augmenting. • Extra arguments can be passed through the implicit parameter list. • .validate() delegates back to the validation object being passed into the method and uses the object being decorated to carry out the validation.
- 18. An Example 18#DevSAIS19 Step 3: Define a validation • Our validation extends a Validation typed with Dataset[Article] • It fills in the abstract method .execute() which defines what the validation is checking for • This means that any time the compiler finds a Dataset[Article] type, we can call .validate() on it with this validation supplied because of our implicit class • Roughly 20 lines of concise and isolated code is nicely separated from the core ETL job
- 19. An Example 19#DevSAIS19 Step 4: Instantiate your validation and pull it in scope • This is what triggers the compiler to link Datasets of Articles to the .validate() method through the defined implicit class
- 20. An Example 20#DevSAIS19 Step 5: Don’t forget Unit Tests • It is straightforward to develop concise and isolated unit tests for each validation that is developed • ScalaTest with FunSpec are used to achieve BDD-style tests
- 21. An Example 21#DevSAIS19 Step 6: And we’re done! • We have been able to develop concise, isolated, testable code that can fit seamlessly into existing Spark jobs • Data is messy, and we have the ability to address this problem in an elegant way • “Enrich my library” has allowed us to extend Spark APIs so we can stay true to functional programming paradigms
- 22. Agenda • Motivation • Scala’s “Enrich My Library” Pattern • An Example • Other Uses 22#DevSAIS19
- 23. Other Uses 23#DevSAIS19 Code organization and readability • Move long blocks of related ETL code into implicit class function definitions to help organize code
- 24. Other Uses 24#DevSAIS19 Support other common functionalities used in production systems ü Validations • Metrics Collection • Logging • Checkpointing • Notifications • …
- 25. Disclaimer These are powerful programming constructs that can greatly increase productivity and enable the buildout of concise and elegant framework code. Overuse can lead to cryptic and esoteric systems that can cause engineers great pain and suffering. Find the right balance! 25#DevSAIS19
- 26. Takeaways • The “Enrich My Library” programming pattern enables concise, clean, and readable code • It enabled us to create a framework that supports rapid development of new validations with a relatively small amount of code • The resulting code is isolated, testable, and easy to understand 26#DevSAIS19
- 27. Come Work At Target • We are hiring in Data Science and Data Engineering • Solve real-world problems ranging from supply chain logistics to smart stores to personalization and so on • Offices in… o Sunnyvale, CA o Minneapolis, MN o Pittsburgh, PA o Bangalore, India 27#DevSAIS19 work somewhere you
- 28. Acknowledgements • Thank you Spark Summit • Thank you Target • Thank you wonderful team members at Target • Thank you vibrant Spark and Scala communities 28#DevSAIS19