Make Structs Easy (MSE)

This library adds withField, withFieldRenamed, and dropFields methods to the Column class allowing users to easily add, rename, and drop fields inside of StructType columns. The signature and behaviour of these methods is intended to be similar to their Dataset equivalents, namely the withColumn, withColumnRenamed, and drop methods.

The methods themselves are backed by efficient Catalyst Expressions and as a result, should provide better performance than equivalent UDFs. While this library uses Scala's implicit conversion technique to "monkey patch" the methods on to the Column class, there is an on-going effort to add these methods natively to the Column class in the Apache Spark SQL project. You can follow along with the progress of this initiative in SPARK-22231.

If you find this project useful, please consider supporting it by giving a star!

Supported Spark versions

MSE should work without any further requirements on Spark/PySpark 2.4.x. The library is available for both Scala versions 2.11 and 2.12. The library is available for Python 3.x.

Installation

Scala

Stable releases of MSE are published to Maven Central. As such, you can pull in the current stable release by simply adding a library dependency to your project for the correct version. For example, for an SBT project, simply add the following line to your build.sbt:

libraryDependencies += "com.github.fqaiser94" %% "mse" % "0.2.4"

For other types of projects (e.g. Maven, Gradle), see the installation instructions at this link.

Python

Stable releases of MSE are published to PyPi. You will also need to provide your PySpark application/s with the path to the MSE jar which you can get from here. For example:

pip install mse
curl https://repo1.maven.org/maven2/com/github/fqaiser94/mse_2.11/0.2.4/mse_2.11-0.2.4.jar --output mse.jar
pyspark --jars mse.jar

If you get errors like TypeError: 'JavaPackage' object is not callable, this usually indicates that you haven't provided PySpark with the correct path to the MSE jar.

Usage

To bring in to scope the (implicit) Column methods in Scala, use:

import com.github.fqaiser94.mse.methods._

To bring in to scope the (implicit) Column methods in Python, use:

from mse import *

The rest of the example code shown below is written in Scala although equivalent Python code would look very similar.

You can now use these methods to manipulate fields in a top-level StructType column:

import org.apache.spark.sql._
import org.apache.spark.sql.types._

// Generate some example data
val structLevel1 = spark.createDataFrame(sc.parallelize(
  Row(Row(1, null, 3)) :: Nil),
  StructType(Seq(
    StructField("a", StructType(Seq(
      StructField("a", IntegerType),
      StructField("b", IntegerType),
      StructField("c", IntegerType))))))).cache
      
structLevel1.show
// +-------+                                                                       
// |      a|
// +-------+
// |[1,, 3]|
// +-------+

structLevel1.printSchema
// root
//  |-- a: struct (nullable = true)
//  |    |-- a: integer (nullable = true)
//  |    |-- b: integer (nullable = true)
//  |    |-- c: integer (nullable = true)

// add new field to top level struct
structLevel1.withColumn("a", 'a.withField("d", lit(4))).show
// +----------+
// |         a|
// +----------+
// |[1,, 3, 4]|
// +----------+

// replace field in top level struct
structLevel1.withColumn("a", 'a.withField("b", lit(2))).show
// +---------+
// |        a|
// +---------+
// |[1, 2, 3]|
// +---------+

// rename field in top level struct
structLevel1.withColumn("a", 'a.withFieldRenamed("b", "z")).printSchema
// root
//  |-- a: struct (nullable = true)
//  |    |-- a: integer (nullable = true)
//  |    |-- z: integer (nullable = true)
//  |    |-- c: integer (nullable = true)

// drop field in top level struct
structLevel1.withColumn("a", 'a.dropFields("b")).show
// +------+
// |     a|
// +------+
// |[1, 3]|
// +------+

You can also use these methods to manipulate fields in nested StructType columns:

// Generate some example data  
val structLevel2 = spark.createDataFrame(sc.parallelize(
    Row(Row(Row(1, null, 3))) :: Nil),
    StructType(Seq(
      StructField("a", StructType(Seq(
        StructField("a", StructType(Seq(
          StructField("a", IntegerType),
          StructField("b", IntegerType),
          StructField("c", IntegerType)))))))))).cache
          
structLevel2.show
// +---------+
// |        a|
// +---------+
// |[[1,, 3]]|
// +---------+

structLevel2.printSchema
// |-- a: struct (nullable = true)
// |    |-- a: struct (nullable = true)
// |    |    |-- a: integer (nullable = true)
// |    |    |-- b: integer (nullable = true)
// |    |    |-- c: integer (nullable = true)

// add new field to nested struct
structLevel2.withColumn("a", 'a.withField(
  "a", $"a.a".withField("d", lit(4)))).show
// +------------+
// |           a|
// +------------+
// |[[1,, 3, 4]]|
// +------------+

// replace field in nested struct
structLevel2.withColumn("a", $"a".withField(
  "a", $"a.a".withField("b", lit(2)))).show
// +-----------+
// |          a|
// +-----------+
// |[[1, 2, 3]]|
// +-----------+
    
// rename field in nested struct
structLevel2.withColumn("a", 'a.withField(
  "a", $"a.a".withFieldRenamed("b", "z"))).printSchema
// |-- a: struct (nullable = true)
// |    |-- a: struct (nullable = true)
// |    |    |-- a: integer (nullable = true)
// |    |    |-- z: integer (nullable = true)
// |    |    |-- c: integer (nullable = true)
    
// drop field in nested struct
structLevel2.withColumn("a", 'a.withField(
  "a", $"a.a".dropFields("b"))).show
// +--------+
// |       a|
// +--------+
// |[[1, 3]]|
// +--------+

You can also manipulate deeply nested StructType columns using the aforementioned patterns but it can be a little annoying to write out the full chain. For this scenario, this library also provides a helper method, namely add_struct_field. You can use this method to add, rename, and drop deeply nested fields as shown below:

// Generate some example data  
val structLevel3 = spark.createDataFrame(sc.parallelize(
    Row(Row(Row(Row(1, null, 3)))) :: Nil),
    StructType(Seq(
      StructField("a", StructType(Seq(
        StructField("a", StructType(Seq(
          StructField("a", StructType(Seq(
              StructField("a", IntegerType),
              StructField("b", IntegerType),
              StructField("c", IntegerType))))))))))))).cache
              
structLevel3.show
//+-----------+
//|          a|
//+-----------+
//|[[[1,, 3]]]|
//+-----------+

structLevel3.printSchema
//root
// |-- a: struct (nullable = true)
// |    |-- a: struct (nullable = true)
// |    |    |-- a: struct (nullable = true)
// |    |    |    |-- a: integer (nullable = true)
// |    |    |    |-- b: integer (nullable = true)
// |    |    |    |-- c: integer (nullable = true)

// add new field to deeply nested struct
structLevel3.withColumn("a", add_struct_field("a.a.a", "d", lit(4))).show
// +--------------+                                                                
// |             a|
// +--------------+
// |[[[1,, 3, 4]]]|
// +--------------+

// replace field in deeply nested struct
structLevel3.withColumn("a", add_struct_field("a.a.a", "b", lit(2))).show
// +-------------+
// |            a|
// +-------------+
// |[[[1, 2, 3]]]|
// +-------------+
    
// rename field in deeply nested struct
structLevel3.withColumn("a", add_struct_field("a.a", "a", $"a.a.a".withFieldRenamed("b", "z"))).printSchema
// root
//  |-- a: struct (nullable = true)
//  |    |-- a: struct (nullable = true)
//  |    |    |-- a: struct (nullable = true)
//  |    |    |    |-- a: integer (nullable = true)
//  |    |    |    |-- z: integer (nullable = true)
//  |    |    |    |-- c: integer (nullable = true)

// drop field in deeply nested struct
structLevel3.withColumn("a", add_struct_field("a.a", "a", $"a.a.a".dropFields("b"))).show
// +----------+
// |         a|
// +----------+
// |[[[1, 3]]]|
// +----------+

// add, rename, and drop fields in deeply nested struct
val result = structLevel3.withColumn("a", add_struct_field("a.a", "a", $"a.a.a".dropFields("b").withFieldRenamed("c", "b").withField("c", lit(4))))
result.show
// +-------------+
// |            a|
// +-------------+
// |[[[1, 3, 4]]]|
// +-------------+

result.printSchema
// root
//  |-- a: struct (nullable = true)
//  |    |-- a: struct (nullable = true)
//  |    |    |-- a: struct (nullable = true)
//  |    |    |    |-- a: integer (nullable = true)
//  |    |    |    |-- b: integer (nullable = true)
//  |    |    |    |-- c: integer (nullable = false)

Another common use-case is to perform these operations on arrays of structs. To do this using the Scala APIs, we recommend combining the functions in this library with the functions provided in spark-hofs:

import org.apache.spark.sql._
import org.apache.spark.sql.types._
import za.co.absa.spark.hofs._
import com.github.fqaiser94.mse.methods._

// Generate some example data
val arrayOfStructs = spark.createDataFrame(sc.parallelize(
    Row(List(Row(1, null, 3), Row(4, null, 6))) :: Nil),
    StructType(Seq(
      StructField("array", ArrayType(
        StructType(Seq(
          StructField("a", IntegerType),
          StructField("b", IntegerType), 
          StructField("c", IntegerType)))))))).cache
          
arrayOfStructs.show
// +------------------+
// |             array|
// +------------------+
// |[[1,, 3], [4,, 6]]|
// +------------------+

arrayOfStructs.printSchema
// root
//  |-- array: array (nullable = true)
//  |    |-- element: struct (containsNull = true)
//  |    |    |-- a: integer (nullable = true)
//  |    |    |-- b: integer (nullable = true)
//  |    |    |-- c: integer (nullable = true)

// add new field to each struct element of array 
arrayOfStructs.withColumn("array", transform($"array", elem => elem.withField("d", lit("hello")))).show(false)
// +--------------------------------+
// |array                           |
// +--------------------------------+
// |[[1,, 3, hello], [4,, 6, hello]]|
// +--------------------------------+

// replace field in each struct element of array
arrayOfStructs.withColumn("array", transform($"array", elem => elem.withField("b", elem.getField("a") + 1))).show(false)
// +----------------------+
// |array                 |
// +----------------------+
// |[[1, 2, 3], [4, 5, 6]]|
// +----------------------+

// rename field in each struct element of array
arrayOfStructs.withColumn("array", transform($"array", elem => elem.withFieldRenamed("b", "z"))).printSchema
// root
//  |-- array: array (nullable = true)
//  |    |-- element: struct (containsNull = true)
//  |    |    |-- a: integer (nullable = true)
//  |    |    |-- z: integer (nullable = true)
//  |    |    |-- c: integer (nullable = true)

// drop field in each Struct element of array
arrayOfStructs.withColumn("array", transform($"array", elem => elem.dropFields("b"))).show(false)
// +----------------+
// |array           |
// +----------------+
// |[[1, 3], [4, 6]]|
// +----------------+

SQL installation and usage

The underlying Catalyst Expressions are SQL compatible. Unfortunately, Spark only added public APIs for plugging in custom Catalyst Expressions into the FunctionRegistry in Spark 3.0.0 (which is at the time of writing is still in preview). You can find a project with an example of how to do this here.

Catalyst Optimization Rules

We also provide some Catalyst optimization rules that can be plugged into a Spark session to get even better performance. This is as simple as including the following two lines of code at the start of your Scala Spark program:

import org.apache.spark.sql.catalyst.optimizer.SimplifyStructExpressions
spark.experimental.extraOptimizations = SimplifyStructExpressions.rules

Spark will use these optimization rules to internally rewrite queries in a more optimal fashion. For example, consider the following query and its corresponding physical plan:

val query = structLevel1.withColumn("a", 'a.withField("d", lit(4)).withField("e", lit(5)))

query.explain
// == Physical Plan ==
// *(1) Project [add_fields(add_fields(a#1, d, 4), e, 5) AS a#32343]
// +- InMemoryTableScan [a#1]
//       +- InMemoryRelation [a#1], StorageLevel(disk, memory, deserialized, 1 replicas)
//             +- Scan ExistingRDD[a#1]

If we add the SimplifyStructExpressions.rules to our Spark session, we see a slightly different physical plan for the same query:

import org.apache.spark.sql.catalyst.optimizer.SimplifyStructExpressions
spark.experimental.extraOptimizations = SimplifyStructExpressions.rules

query.explain
// == Physical Plan ==
// *(1) Project [add_fields(a#1, d, e, 4, 5) AS a#32343]
// +- InMemoryTableScan [a#1]
//       +- InMemoryRelation [a#1], StorageLevel(disk, memory, deserialized, 1 replicas)
//             +- Scan ExistingRDD[a#1]

As you can see, the successive add_fields method calls have been collapsed into a single add_fields method call.

Theoretically, this should improve performance but for the most part, you won't notice much difference unless you're doing some particularly intense struct manipulation and/or working with a particularly large dataset.

Unfortunately, to the best of our knowledge, there is currently no way to plug in custom Catalyst optimization rules directly using the Python APIs.

Questions/Thoughts/Concerns?

Feel free to submit an issue.

Instructions for deploying a new release

Increment version number in python/setup.py
Create a new release with appropriately incremented tag

fqaiser94 / mse 0.2.4