Explore optimization techniques for complex Apache Spark plans in this 27-minute conference talk from DAIS NA 2021. Dive into Workday's experience building analytics products with Spark, addressing challenges like compiling large-scale DataFrames and handling extensive case statements. Learn about memory-efficient plan logging, common subexpression elimination for redundant subplan removal, and rewriting Spark's constraint propagation mechanism. Discover how these enhancements improve Catalyst performance on production pipelines, and gain valuable tips for managing complex Spark plans in your own projects. The talk covers topics such as data validation, handling large case expressions, optimized constraint propagation, and future improvements in Spark optimization.
Overview
Syllabus
Intro
Spark in Workday Prism Analytics
Example: Data Validation
About Complex Plans
Common Subexpression Elimination (CSE)
CSE Benchmark
Logging Complex Plans (10s of MBs in Size)
Problems with Large Case Expressions
Handling Large Case Expressions in Catalyst
Large Case Expression Benchmark
Example: Generate New Filter
Example: Prune Redundant Filter
Example: New Filter on Other Side of Join
Current Constraint Propagation Algorithm
Current Algorithm Takes High Memory
Recall: Fix for Large Case Expressions
Optimized Constraint Propagation (SPARK-33152)
Constraint Propagation Algorithms Comparison
Constraint Propagation Benchmark
Effect on Customer Pipeline
Tuning Tips
Future Work
Taught by
Databricks
