zero323 / spark-gotchas

Spark Gotchas. A subjective compilation of the Apache Spark tips and tricks

Spark Gotchas

Table of Contents

• Introduction
  • Version compatibility
• Understanding Spark Architecture
  • Lack of Global Shared State and Spark Closures
  • Distributed Processing and its Scope
  • Feature Parity and Architecture of the Guest Languages
• Spark Application Deployment
  • Dynamic allocation
    • Yet Another Dynamic Resource Allocation YA(D)RN
• Spark Application Building
• RDD actions and Transformations by Example
  • Be Smart About groupByKey
    • What Exactly Is Wrong With groupByKey
    • How Not to Optimize
    • Not All groupBy Methods Are Equal
    • When to Use groupByKey and When to Avoid It
    • Hidden groupByKey
  • Immutability of a Data Structure Does Not Imply Immutability of the Data
• Spark SQL and Dataset API
  • Non-lazy Evaluation in DataFrameReader.load
    • Explicit Schema
    • Sampling for Schema Inference
    • PySpark Specific Considerations
  • DataFrame Schema Nullablility
    • nullability by Reflection
    • Marking StructFields Excplicitly as Nullable
    • Impact of Nullable
  • Reading Data Using JDBC Source
    • Parallelizing Reads
    • MySQL: Dates, Timestamps and Lies
  • Window Functions
    • Understanding Window Functions
    • Window Definitions
    • Example Usage
    • Requirements and Performance Considerations
• Data Preparation
  • DataFrame Metadata
    • Metadata in ML pipelines
    • Setting custom column metadata
• Iterative Algorithms
  • Iterative Applications and Lineage
    • Checkpointing
    • "Flat Transformations"
    • Truncating Lineage in Dataset API
  • Controling Number of Partitions in Iterative Applications
• PySpark Applications
• Serialization
  • JVM
    • Java Serialization
    • Kryo Serialization
  • PySpark Serialization
    • Python Serializers Characteristics
    • PySpark Serialization Strategies
    • Configuring PySpark Serialization
    • PySpark and Kryo
  • SerDe During JVM - Guest Communication

License

This work, excluding code examples, is licensed under Creative Commons Attribution-ShareAlike 4.0 International license.

Accompanying code and code snippets are licensed under MIT license.