How Not To Persist Spark Dataset / DataFrame toNeo4j (And How To?) 1

Spark provide an API for graphs and graph-parallel computation (GraphX) and Neo4j is probably the most popular graph database so you might expect some sort of “bridge” between them. That’s actually the case, and Neo4j will direct you to the Neo4j-Spark-Connector. It’s a Scala based implementation that uses the officially supported Neo4j Java driver behind the hood. It covers a series of use cases but, I argue, not the one mentioned in the title.

The scenario I’m considering could result from processing data with GraphX and then persisting the results in Neo4j where they could then be read by other process (including another Spark job).  Probably because is not such an unusual use case, there are people asking details about it on  Stackoverflow and in this issue.

Neo4j-Spark-Connector allegedly cover this case and there is also an example in: Neo4jDataFrame.mergeEdgeList. If we take a look at the code though, mergeEdgeList basically map over the DataFrame and for each row call a method called execute:

def mergeEdgeList(sc: SparkContext,
                      dataFrame: DataFrame,
                      source: (String,Seq[String]),
                      relationship: (String,Seq[String]),
                      target: (String,Seq[String]),
                      renamedColumns: Map[String,String] = Map.empty): Unit = {
   val partitions = Math.max(1,(dataFrame.count() / 10000).asInstanceOf[Int])
   val config = Neo4jConfig(sc.getConf)
   dataFrame.repartition(partitions).foreachPartition( rows => {
      ...
      execute(config, mergeStatement, Map("rows" -> params).asJava, write = true)
   })
}

And then execute will do the proper writing in the db:

def execute(config : Neo4jConfig, query: String, parameters: java.util.Map[String, AnyRef], write: Boolean = false) : ResultSummary = {
    val driver: Driver = config.driver()
    val session = driver.session()
    try {
      val runner =  new TransactionWork[ResultSummary]() {
         override def execute(tx:Transaction) : ResultSummary =
            tx.run(query, parameters).consume()
      }
      if (write) {
        session.writeTransaction(runner)
      }
      else
        session.readTransaction(runner)
    } finally {
      if (session.isOpen) session.close()
      driver.close()
    }
  }

There are at least three issues with the code above. Let’s start with the first one:

1. In mergeEdgeList, count and foreachPartition are actions and as such will trigger the execution of the given dataframe twice. And that without the caller being aware of it.
2. execute create a neo4j driver (through the java driver) and, from it, a session, then it closes both. And this is done for each partition in the dataframe.
3. Because of 2. you’ll have a session per partition hence the scope of the transaction will span over a single partition. Now, what the scope of a transaction should be, has to be decided case by case but you won’t have that freedom of choice here. For example you might need a broader scope (fail or succeed to upload the full dataframe so there are only two possible outcomes: either the dataframe has been written correctly in Neo4j or it hasn’t) or no transaction support at all.

Point 1 should be addressed as it could be pretty insidious or at least made the caller aware that the dataframe needs to be cached. Overall the Neo4j-Spark-Connector seems to fit other scenarios (streaming, reading, update, etc) but probably shouldn’t be the first pick when a potentially demanding spark job is meant to create a new Neo4j db from scratch.

After some research, I reckon the best way (also from a performance point of view) is to persist the dataframe as csv and leverage the Neo4j load csv functionality. Although it might look a pretty clumsy solution at first, the csv header can be tailored in order to pass some useful information to the Neo4j csv loader. For example the csv headers:

movieId:ID,title,year:int,:LABEL

Tell the loader that movieId is the node’s id (and its uniqueness will be enforced), year is a node’s property of type int and the value under the third column in the csv is a label for the node.

I wrote a simple proof of concept that illustrate the Neo4j csv loader functionalities: https://github.com/lansaloltd/spark-neo4j

The example is fictional and loosely based on a de-duplication problem. The starting point is a dataset where couple of documents were marked as duplicate by a machine learning algorithm. The data is filtered first to keep only duplicates and the AI algorithm confidence level becomes a property of the relationship between two documents (how likely they are to be indeed duplicates). In this scenario the output has to be a disconnected graph (as we can’t expect all documents to resemble each other) but with islands or clusters of documents linked one to the others or, more precisely, connected components.

It requires Neo4j Desktop installed (download here). Once started you should see something similar to this:

Click on “Add Graph”:

Then chose a name and a password for your DB (if you like the code example to work out of the box, then chose “password” as password and version 3.5.x):

Click on “Manage” and then “Settings” tab and comment out dbms.directories.import=import then press “Apply”:

That’s because the unit test will create the csv files to import under the target directory which is an external directory for Neo4j. In a production environment that has security implications you should be aware of.

Finally, click “Start” to start the database and once active, you can run the tests in GraphAnalysisTest. If successful you might still see 0 nodes / 0 relationships on the DB section but that seems to be only a refresh issue. Click on Neo4J Browser:

Then on the DB icon on the top left corner:

And now you should see the graph being successfully imported into Neo4j:

There are 31 nodes (or vertexes in graph terminology) and 29 relationships (or edges).

———————————

^{1. [This is an updated version of the original post to amend a mistake in the first version. It was spotted while talking with few colleagues, weeks after the initial publication when still working on the “csv” solution for a presentation: the connection gets created at the foreachPartition level, which means one connection per partition not per record (as erroneously stated).
I apologize for all the time it took me to get back on this post and find the time to update it but these have been quite hectic times (not just from a work perspective).]↩}

Spark ML Model Tuning

The purpose of this post is to provide some basic guidelines and supporting code for the tuning of an ML pipeline based on a random forest algorithm using cross-validation along with some final consideration about the computational cost involved and possibility to train/tuning the model in parallel.

Tuning

A random forest algorithm takes few hyperparameters. The most important ones are probably numTrees and maxDepth (but in code example there is also some reference to maxBins). The number of trees in the forest (numTrees) is inversely proportional to the variance in predictions: more trees generally improve the accuracy but also the computational load of the training phase increase proportionally. Analogously, a deep tree generally lead to better metrics but also increase the risk of overfitting and the training computational cost. Tuning the model is the process of working out the optimal values for those hyperparameters and striking the right balance between costs and benefits.

Cross Validation

The idea is to split a given labelled dataset in chunks (folds) then iterates through them and, at each iteration, one chunk will be used as a validation set and the remaining as a training set. For example, with fold = 3, the process splits the data in three folds: a, b, c and then iterates three times:

1st iteration -> a + b = training, c = validation
2nd iteration -> a + c = training, b = validation
3rd iteration -> b + c = training, a = validation

To note that the whole process is indeed training and evaluating three times (as in this example folds = 3).

An Estimator can then be plugged in the cross validation process with the intention of keeping track of the relationship between certain hyperparameters values and some metrics and pick up the best performing model.

Data and Code

Code can be found here. Inspired by the Cambridge Analytica scandal, and as a warning that great power comes with great responsibility, the example uses a completely fictional and auto generated dataset where made up citizens are listed with the party they voted for (the label) along with some features (the usual suspects: age, income, gender, etc.):

+----+------+---+----+------+---------+-------+
|id  |gender|age|area|income|education|vote   |
+----+------+---+----+------+---------+-------+
|650 |female|99 |A   |57000 |1        |tory   |
|1523|male  |24 |B   |5500  |3        |labour |
|434 |male  |82 |D   |5500  |3        |liberal|
|174 |male  |26 |C   |69000 |1        |liberal|
+----+------+---+----+------+---------+-------+

Road Map

As mentioned, the data in the example is not only fictional but some statistical correlation has been artificially introduced in the dataset and the features engineering part is simplified but on the other hand that is all what is needed in order to illustrate the advantages of tuning. The focus need to be on the different outcome between a tuned versus a not tuned model. Quite obviously, a tuned model is expected to perform better than a non tuned one and the following road map should prove the point:

1. Given a training dataset, train the model without any tuning
2. Using the same training dataset, train the model using cross validation and an evaluator to work out the optimal hyperparameters values
3. Train a new model using the optimal hyperparameters values and same training dataset at point 1
4. Evaluate the model at point 1 (not tuned) and model at point 3 (tuned) against a given validation set and compare the results

Evaluation will be based on f1 metric (which takes into consideration precision and recall). Besides being the default, f1 seems the most natural choice for the example.

Tuning the Model

First point in the road map is to train a random forest model without any specific tuning (ie. default value for numTrees and maxDepth will be used)¹. In the example that is done in a training method that takes as a parameter a given training set.

val vectorAssembler = new VectorAssembler().setInputCols(Array("f_gender", "age", "f_area", "income", "f_education"))
    .setOutputCol("features")
val genderIndexer = new StringIndexer().setInputCol("gender").setOutputCol("f_gender")
val areaIndexer = new StringIndexer().setInputCol("area").setOutputCol("f_area")
val educationIndexer = new StringIndexer().setInputCol("education").setOutputCol("f_education")
val voteIntentionIndexer = new StringIndexer().setInputCol("vote").setOutputCol("label")

def training(trainingSet: DataFrame): PipelineModel = {
  val transformationPipeline = new Pipeline()
  val classifier = new RandomForestClassifier().setSeed(5043) // set the seed for reproducibility
  val trainingPipeline: transformationPipeline.type = transformationPipeline.setStages(
    Array(genderIndexer, areaIndexer, educationIndexer, voteIntentionIndexer, vectorAssembler, classifier))
  trainingPipeline.fit(trainingSet)
}

Then the model can be evaluated against a given validation set and its performance expressed in terms of f1:

def validation(model: PipelineModel, validation: DataFrame) = {
    // Make predictions.
    val predictions = model.transform(validation)

    val evaluator = new MulticlassClassificationEvaluator()
      .setLabelCol("label")
      .setPredictionCol("prediction")
      .setMetricName("f1") // f1 is the default anyway

    val f1: Double = evaluator.evaluate(predictions)
    println("@@ f1 ====> " + f1)
  }

The whole flow is implemented in a test in VoteIntentionTrainingTest called train and evaluate model. Once executed it should return a value for f1 of 0.6736260921544567 and this will the benchmark to compare the tuned version against.

Better Hyperparameters Values

Figuring out the optimal hyperparameters value should be more interesting. A ParamGridBuilder allows to pass different candidates for each hyperparameter. Considering the Spark default values (numTrees = 20, maxDepth = 5), a good first approximation would be two values surrounding the default. For example numTrees is 20 hence passing an Array(15, 20, 25) will give an idea about whether better metrics can be obtained with values greater or smaller than the default one.

def tuning(trainingSet: DataFrame): CrossValidatorModel = {

  val rfClassifier = new RandomForestClassifier()
  val pipeline= new Pipeline().setStages(
    Array(genderIndexer, areaIndexer, educationIndexer, voteIntentionIndexer, vectorAssembler, rfClassifier))

  val nFolds: Int = 8
  val metric: String = "f1"
  val paramGrid = new ParamGridBuilder()
    .addGrid(rfClassifier.numTrees, Array(15, 20, 25))
    .addGrid(rfClassifier.maxDepth, Array(4, 5, 6))
    .build()
  val evaluator = new MulticlassClassificationEvaluator() .setLabelCol("label") .setPredictionCol("prediction") .setMetricName(metric)

  val cv: CrossValidator = new CrossValidator()
    .setEstimator(pipeline)
    .setEstimatorParamMaps(paramGrid)
    .setEvaluator(evaluator)
    .setNumFolds(nFolds)
    .setParallelism(3)

  val fittedPipeline: CrossValidatorModel = cv.fit(trainingSet)
  val best: Model[_] = fittedPipeline.bestModel
  val stages: Seq[Transformer] = best.asInstanceOf[PipelineModel].stages.toList


  val paramMap = fittedPipeline.getEstimatorParamMaps
    .zip(fittedPipeline.avgMetrics)
    .maxBy(_._2)
    ._1

  println(s"@@@@ best num trees:     ${stages(5).asInstanceOf[RandomForestClassificationModel].getNumTrees}")
  println(s"@@@@ best max depth:     ${stages(5).asInstanceOf[RandomForestClassificationModel].getMaxDepth}")
  println(s"@@@@ best max bins:      ${stages(5).asInstanceOf[RandomForestClassificationModel].getMaxBins}")

  println(s"@@@@ best paramMap:      ${paramMap}")

  fittedPipeline
}

Once the pipeline is fitted we grab the best model, find out the optimal hyperparameters values (among the ones provided in ParamGridBuilder) and print them in the console. As a note, we are interested in stage 6 (so the 5th element in the array of stages) because the pipeline (as defined in this example) has 6 stages, the last one being the classifier (the one we are interested in) but paramMap returns the same results. The whole process is implemented in VoteIntentionTrainingTest as the test tuning a model and save it in $TunedModelPath and if executed should return the following optimal values:

num trees:     25
max depth:     6

We could refine those results further and for each hyperparameter pass new candidates in an interval around the optimal value. For example, in the numTrees case an Array(23, 24, 25, 26, 27) could be passed to the ParamGridBuilder with a new set candidates to fulfil the role of optimal value. I’ll omit that here but that should return the following values:

num trees:     24
max depth:     6

Tuned vs Not Tuned Model

The optimal values can now be used to override the default ones:

  def training(trainingSet: DataFrame): PipelineModel = {
    val transformationPipeline = new Pipeline()
    val classifier = new RandomForestClassifier()
      .setNumTrees(24) // <- Tuned value
      .setMaxDepth(6) // <- Tuned value
      .setSeed(5043) // set the seed for reproducibility
    val trainingPipeline: transformationPipeline.type = transformationPipeline.setStages(
      Array(genderIndexer, areaIndexer, educationIndexer, voteIntentionIndexer, vectorAssembler, classifier))
    trainingPipeline.fit(trainingSet)
  }

And an evaluation of the new model against the same validation dataset, should return an f1 of 0.6919821403245205 which is an improvement if compared to the previous value of f1 (0.6736260921544567) obtained with the default hyperparameters value.

numTrees = 20, maxDepth = 5 -> f1 = 0.6736260921544567 (default values)
numTrees = 24, maxDepth = 6 -> f1 = 0.6919821403245205

Training and Parallelism

It is worth noting that cross validation and the parameter grid increase substantially the computational cost of fitting a model. The first tuning example:

.addGrid(rfClassifier.numTrees, Array(15, 20, 25))
.addGrid(rfClassifier.maxDepth, Array(4, 5, 6))

will evaluate 9 different parameter combinations (3 * 3) and this on the top of the cross validation cost (8 folds). In this fictional example we are dealing with dataset intentionally small but in a real case scenario Spark and a distributed architecture are likely to make a big difference and we can take advantage of this via the setParallelism method.

val cv: CrossValidator = new CrossValidator()
      .setEstimator(pipeline)
      .setEstimatorParamMaps(paramGrid)
      .setEvaluator(evaluator)
      .setNumFolds(nFolds)
      .setParallelism(3)

That would set the max number of threads to use when running parallel algorithms (default is 1 for serial execution).

———————————

^{1. [In Spark 2.4.3 the default values are: numTrees = 20, maxDepth = 5, maxBins = 32]↩}

Spark: How to Add Multiple Columns in Dataframes (and How Not to)

There are generally two ways to dynamically add columns to a dataframe in Spark. A foldLeft or a map (passing a RowEncoder). The foldLeft way is quite popular (and elegant) but recently I came across an issue regarding its performance when the number of columns to add is not trivial. I think it’s worth to share the lesson learned: a map solution offers substantial better performance when the number of columns to be added tends to increase.

Let’s start with an example dataframe with four columns: “id”, “author”, “title”, “incipit” (the opening line of the novel). A simple dataframe with just one row might look something like:

+---+-----------+-------------+-----------------------------+
| id|     author|        title|                      incipit|
+---+-----------+-------------+-----------------------------+
|  6|Leo Tolstoy|Anna Karenina|Happy families are all alike |
+---+-----------+-------------+-----------------------------+

You are given an arbitrary list of words and, for each of them, you would like to add a column (named after the word) to the original dataframe and flag with a boolean whether or not that word appear at least once in the opening line (incipit).
For a list of two words List("families", "giraffe"), the above dataframe will be transformed into the following:

+---+-----------+-------------+-----------------------------+--------+-------+
| id|     author|        title|                      incipit|families|giraffe|
+---+-----------+-------------+-----------------------------+--------+-------+
|  6|Leo Tolstoy|Anna Karenina|Happy families are all alike |    true|  false|
+---+-----------+-------------+-----------------------------+--------+-------+

As the list of columns is arbitrary, there are two possible approaches to this problem. I wrapper both in a method to make testing easier. First approach would be the foldLeft way:

def addColumnsViaFold(df: DataFrame, columns: List[String]): DataFrame = {
  import df.sparkSession.implicits._
  columns.foldLeft(df)((acc, col) => {
    acc.withColumn(col, acc("incipit").as[String].contains(col))
  })
}

And the second one (which involves a bit more coding) is the map way:

def addColumnsViaMap(df: DataFrame, words: List[String]): DataFrame = {
   val encoder = RowEncoder.apply(getSchema(df, words))
   df.map(mappingRows(df.schema)(words))(encoder)
}

private val mappingRows: StructType => List[String] => Row => Row =
  (schema) => (words) => (row) => {
    val addedCols: List[Boolean] = words.map {
      word => row.getString(schema.fieldIndex("incipit")).contains(word)
    }
    Row.merge(row, Row.fromSeq(addedCols))
  }

private def getSchema(df: DataFrame, words: List[String]): StructType = {
  var schema: StructType = df.schema
  words.foreach(word => schema = schema.add(word, DataTypes.BooleanType, false))
  schema
}

The code (including all tests) is available here
When we run the scala meter tests to get some idea of how the two approaches behave when dealing with 100 new columns, we get the following results¹. For foldLeft (addColumnsViaFold method):

Whereas those one are the results for map (addColumnsViaMap method):

When the number of columns increases, foldLeft is taking considerably more time. If we take the number of columns to 500 the result is similar (and more dramatic)
Once again, for foldLeft :

And for map:

With 1000 columns, foldLeft job aborts:

[error] Error running separate JVM: java.lang.OutOfMemoryError: GC overhead limit exceeded

Probably sign that the heap is running low and the CG can’t free much memory. Still, the map based solution seems to cope much better even with 1000 row:

When it comes to the reason behind this different behavior, my guess would be that somehow Catalyst is not able to optimize the foldLeft operation. The below plans (explain(true)) show how the Parsed and Analyzed Logical Plans end up in a sort of nested structure.

== Parsed Logical Plan ==
Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 20 more fields]
+- AnalysisBarrier
      +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 19 more fields]
         +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 18 more fields]
            +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 17 more fields]
               +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 16 more fields]
                  +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 15 more fields]
                     +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 14 more fields]
                        +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 13 more fields]
                           +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 12 more fields]
                              +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 11 more fields]
                                 +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 10 more fields]
                                    +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 9 more fields]
                                       +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 8 more fields]
                                          +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 7 more fields]
                                             +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 6 more fields]
                                                +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 5 more fields]
                                                   +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 4 more fields]
                                                      +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 3 more fields]
                                                         +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 2 more fields]
                                                            +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, Contains(incipit#12, toing) AS toing#368]
                                                               +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, Contains(incipit#12, hence) AS hence#342]
                                                                  +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, Contains(incipit#12, four) AS four#317]
                                                                     +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, Contains(incipit#12, John) AS John#293]
                                                                        +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, Contains(incipit#12, pop) AS pop#270]
                                                                           +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, Contains(incipit#12, drunk) AS drunk#248]
                                                                              +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, Contains(incipit#12, two) AS two#227]
                                                                                 +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, Contains(incipit#12, when) AS when#207]
                                                                                    +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, Contains(incipit#12, city) AS city#188]
                                                                                       +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, Contains(incipit#12, morning) AS morning#170]
                                                                                          +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, Contains(incipit#12, way) AS way#153]
                                                                                             +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, Contains(incipit#12, pig) AS pig#137]
                                                                                                +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, Contains(incipit#12, happy) AS happy#122]
                                                                                                   +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, Contains(incipit#12, was) AS was#108]
                                                                                                      +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, Contains(incipit#12, anna) AS anna#95]
                                                                                                         +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, Contains(incipit#12, in) AS in#83]
                                                                                                            +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, Contains(incipit#12, man) AS man#72]
                                                                                                               +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, Contains(incipit#12, with) AS with#62]
                                                                                                                  +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, Contains(incipit#12, alike) AS alike#53]
                                                                                                                     +- Project [id#9, author#10, title#11, incipit#12, to#38, Contains(incipit#12, get) AS get#45]
                                                                                                                        +- Project [id#9, author#10, title#11, incipit#12, Contains(incipit#12, to) AS to#38]
                                                                                                                           +- Project [_1#4 AS id#9, _2#5 AS author#10, _3#6 AS title#11, _4#7 AS incipit#12]
                                                                                                                              +- LocalRelation [_1#4, _2#5, _3#6, _4#7]

== Analyzed Logical Plan ==
id: int, author: string, title: string, incipit: string, to: boolean, get: boolean, alike: boolean, with: boolean, man: boolean, in: boolean, anna: boolean, was: boolean, happy: boolean, pig: boolean, way: boolean, morning: boolean, city: boolean, when: boolean, two: boolean, drunk: boolean, pop: boolean, John: boolean, four: boolean, hence: boolean, ... 20 more fields
Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 20 more fields]
+- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 19 more fields]
   +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 18 more fields]
      +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 17 more fields]
         +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 16 more fields]
            +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 15 more fields]
               +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 14 more fields]
                  +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 13 more fields]
                     +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 12 more fields]
                        +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 11 more fields]
                           +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 10 more fields]
                              +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 9 more fields]
                                 +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 8 more fields]
                                    +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 7 more fields]
                                       +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 6 more fields]
                                          +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 5 more fields]
                                             +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 4 more fields]
                                                +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 3 more fields]
                                                   +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, ... 2 more fields]
                                                      +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, hence#342, Contains(incipit#12, toing) AS toing#368]
                                                         +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, four#317, Contains(incipit#12, hence) AS hence#342]
                                                            +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, John#293, Contains(incipit#12, four) AS four#317]
                                                               +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, pop#270, Contains(incipit#12, John) AS John#293]
                                                                  +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, drunk#248, Contains(incipit#12, pop) AS pop#270]
                                                                     +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, two#227, Contains(incipit#12, drunk) AS drunk#248]
                                                                        +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, when#207, Contains(incipit#12, two) AS two#227]
                                                                           +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, city#188, Contains(incipit#12, when) AS when#207]
                                                                              +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, morning#170, Contains(incipit#12, city) AS city#188]
                                                                                 +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, way#153, Contains(incipit#12, morning) AS morning#170]
                                                                                    +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, pig#137, Contains(incipit#12, way) AS way#153]
                                                                                       +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, happy#122, Contains(incipit#12, pig) AS pig#137]
                                                                                          +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, was#108, Contains(incipit#12, happy) AS happy#122]
                                                                                             +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, anna#95, Contains(incipit#12, was) AS was#108]
                                                                                                +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, in#83, Contains(incipit#12, anna) AS anna#95]
                                                                                                   +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, man#72, Contains(incipit#12, in) AS in#83]
                                                                                                      +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, with#62, Contains(incipit#12, man) AS man#72]
                                                                                                         +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, alike#53, Contains(incipit#12, with) AS with#62]
                                                                                                            +- Project [id#9, author#10, title#11, incipit#12, to#38, get#45, Contains(incipit#12, alike) AS alike#53]
                                                                                                               +- Project [id#9, author#10, title#11, incipit#12, to#38, Contains(incipit#12, get) AS get#45]
                                                                                                                  +- Project [id#9, author#10, title#11, incipit#12, Contains(incipit#12, to) AS to#38]
                                                                                                                     +- Project [_1#4 AS id#9, _2#5 AS author#10, _3#6 AS title#11, _4#7 AS incipit#12]
                                                                                                                        +- LocalRelation [_1#4, _2#5, _3#6, _4#7]

== Optimized Logical Plan ==
Project [id#9, author#10, title#11, incipit#12, Contains(incipit#12, to) AS to#38, Contains(incipit#12, get) AS get#45, Contains(incipit#12, alike) AS alike#53, Contains(incipit#12, with) AS with#62, Contains(incipit#12, man) AS man#72, Contains(incipit#12, in) AS in#83, Contains(incipit#12, anna) AS anna#95, Contains(incipit#12, was) AS was#108, Contains(incipit#12, happy) AS happy#122, Contains(incipit#12, pig) AS pig#137, Contains(incipit#12, way) AS way#153, Contains(incipit#12, morning) AS morning#170, Contains(incipit#12, city) AS city#188, Contains(incipit#12, when) AS when#207, Contains(incipit#12, two) AS two#227, Contains(incipit#12, drunk) AS drunk#248, Contains(incipit#12, pop) AS pop#270, Contains(incipit#12, John) AS John#293, Contains(incipit#12, four) AS four#317, Contains(incipit#12, hence) AS hence#342, ... 20 more fields]
+- InMemoryRelation [id#9, author#10, title#11, incipit#12], true, 10000, StorageLevel(disk, memory, deserialized, 1 replicas)
      +- LocalTableScan [id#9, author#10, title#11, incipit#12]

== Physical Plan ==
*(1) Project [id#9, author#10, title#11, incipit#12, Contains(incipit#12, to) AS to#38, Contains(incipit#12, get) AS get#45, Contains(incipit#12, alike) AS alike#53, Contains(incipit#12, with) AS with#62, Contains(incipit#12, man) AS man#72, Contains(incipit#12, in) AS in#83, Contains(incipit#12, anna) AS anna#95, Contains(incipit#12, was) AS was#108, Contains(incipit#12, happy) AS happy#122, Contains(incipit#12, pig) AS pig#137, Contains(incipit#12, way) AS way#153, Contains(incipit#12, morning) AS morning#170, Contains(incipit#12, city) AS city#188, Contains(incipit#12, when) AS when#207, Contains(incipit#12, two) AS two#227, Contains(incipit#12, drunk) AS drunk#248, Contains(incipit#12, pop) AS pop#270, Contains(incipit#12, John) AS John#293, Contains(incipit#12, four) AS four#317, Contains(incipit#12, hence) AS hence#342, ... 20 more fields]
+- InMemoryTableScan [author#10, id#9, incipit#12, title#11]
      +- InMemoryRelation [id#9, author#10, title#11, incipit#12], true, 10000, StorageLevel(disk, memory, deserialized, 1 replicas)
            +- LocalTableScan [id#9, author#10, title#11, incipit#12]

Not sure whether that is somehow related to the Catalyst inability to optimize foldLeft but the explained map plan doesn’t show the same nested structure:

== Parsed Logical Plan ==
'SerializeFromObject [validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 0, id), IntegerType) AS id#83, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 1, author), StringType), true, false) AS author#84, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 2, title), StringType), true, false) AS title#85, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 3, incipit), StringType), true, false) AS incipit#86, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 4, to), BooleanType) AS to#87, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 5, get), BooleanType) AS get#88, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 6, alike), BooleanType) AS alike#89, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 7, with), BooleanType) AS with#90, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 8, man), BooleanType) AS man#91, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 9, in), BooleanType) AS in#92, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 10, anna), BooleanType) AS anna#93, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 11, was), BooleanType) AS was#94, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 12, happy), BooleanType) AS happy#95, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 13, pig), BooleanType) AS pig#96, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 14, way), BooleanType) AS way#97, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 15, morning), BooleanType) AS morning#98, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 16, city), BooleanType) AS city#99, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 17, when), BooleanType) AS when#100, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 18, two), BooleanType) AS two#101, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 19, drunk), BooleanType) AS drunk#102, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 20, pop), BooleanType) AS pop#103, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 21, John), BooleanType) AS John#104, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 22, four), BooleanType) AS four#105, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 23, hence), BooleanType) AS hence#106, ... 20 more fields]
+- 'MapElements , interface org.apache.spark.sql.Row, [StructField(id,IntegerType,false), StructField(author,StringType,true), StructField(title,StringType,true), StructField(incipit,StringType,true)], obj#82: org.apache.spark.sql.Row
   +- 'DeserializeToObject unresolveddeserializer(createexternalrow(getcolumnbyordinal(0, IntegerType), getcolumnbyordinal(1, StringType).toString, getcolumnbyordinal(2, StringType).toString, getcolumnbyordinal(3, StringType).toString, StructField(id,IntegerType,false), StructField(author,StringType,true), StructField(title,StringType,true), StructField(incipit,StringType,true))), obj#81: org.apache.spark.sql.Row
      +- AnalysisBarrier
            +- Project [_1#4 AS id#9, _2#5 AS author#10, _3#6 AS title#11, _4#7 AS incipit#12]
               +- LocalRelation [_1#4, _2#5, _3#6, _4#7]

== Analyzed Logical Plan ==
id: int, author: string, title: string, incipit: string, to: boolean, get: boolean, alike: boolean, with: boolean, man: boolean, in: boolean, anna: boolean, was: boolean, happy: boolean, pig: boolean, way: boolean, morning: boolean, city: boolean, when: boolean, two: boolean, drunk: boolean, pop: boolean, John: boolean, four: boolean, hence: boolean, ... 20 more fields
SerializeFromObject [validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 0, id), IntegerType) AS id#83, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 1, author), StringType), true, false) AS author#84, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 2, title), StringType), true, false) AS title#85, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 3, incipit), StringType), true, false) AS incipit#86, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 4, to), BooleanType) AS to#87, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 5, get), BooleanType) AS get#88, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 6, alike), BooleanType) AS alike#89, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 7, with), BooleanType) AS with#90, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 8, man), BooleanType) AS man#91, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 9, in), BooleanType) AS in#92, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 10, anna), BooleanType) AS anna#93, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 11, was), BooleanType) AS was#94, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 12, happy), BooleanType) AS happy#95, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 13, pig), BooleanType) AS pig#96, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 14, way), BooleanType) AS way#97, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 15, morning), BooleanType) AS morning#98, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 16, city), BooleanType) AS city#99, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 17, when), BooleanType) AS when#100, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 18, two), BooleanType) AS two#101, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 19, drunk), BooleanType) AS drunk#102, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 20, pop), BooleanType) AS pop#103, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 21, John), BooleanType) AS John#104, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 22, four), BooleanType) AS four#105, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 23, hence), BooleanType) AS hence#106, ... 20 more fields]
+- MapElements , interface org.apache.spark.sql.Row, [StructField(id,IntegerType,false), StructField(author,StringType,true), StructField(title,StringType,true), StructField(incipit,StringType,true)], obj#82: org.apache.spark.sql.Row
   +- DeserializeToObject createexternalrow(id#9, author#10.toString, title#11.toString, incipit#12.toString, StructField(id,IntegerType,false), StructField(author,StringType,true), StructField(title,StringType,true), StructField(incipit,StringType,true)), obj#81: org.apache.spark.sql.Row
      +- Project [_1#4 AS id#9, _2#5 AS author#10, _3#6 AS title#11, _4#7 AS incipit#12]
         +- LocalRelation [_1#4, _2#5, _3#6, _4#7]

== Optimized Logical Plan ==
SerializeFromObject [validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 0, id), IntegerType) AS id#83, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 1, author), StringType), true, false) AS author#84, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 2, title), StringType), true, false) AS title#85, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 3, incipit), StringType), true, false) AS incipit#86, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 4, to), BooleanType) AS to#87, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 5, get), BooleanType) AS get#88, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 6, alike), BooleanType) AS alike#89, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 7, with), BooleanType) AS with#90, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 8, man), BooleanType) AS man#91, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 9, in), BooleanType) AS in#92, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 10, anna), BooleanType) AS anna#93, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 11, was), BooleanType) AS was#94, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 12, happy), BooleanType) AS happy#95, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 13, pig), BooleanType) AS pig#96, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 14, way), BooleanType) AS way#97, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 15, morning), BooleanType) AS morning#98, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 16, city), BooleanType) AS city#99, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 17, when), BooleanType) AS when#100, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 18, two), BooleanType) AS two#101, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 19, drunk), BooleanType) AS drunk#102, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 20, pop), BooleanType) AS pop#103, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 21, John), BooleanType) AS John#104, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 22, four), BooleanType) AS four#105, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 23, hence), BooleanType) AS hence#106, ... 20 more fields]
+- MapElements , interface org.apache.spark.sql.Row, [StructField(id,IntegerType,false), StructField(author,StringType,true), StructField(title,StringType,true), StructField(incipit,StringType,true)], obj#82: org.apache.spark.sql.Row
   +- DeserializeToObject createexternalrow(id#9, author#10.toString, title#11.toString, incipit#12.toString, StructField(id,IntegerType,false), StructField(author,StringType,true), StructField(title,StringType,true), StructField(incipit,StringType,true)), obj#81: org.apache.spark.sql.Row
      +- InMemoryRelation [id#9, author#10, title#11, incipit#12], true, 10000, StorageLevel(disk, memory, deserialized, 1 replicas)
            +- LocalTableScan [id#9, author#10, title#11, incipit#12]

== Physical Plan ==
*(1) SerializeFromObject [validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 0, id), IntegerType) AS id#83, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 1, author), StringType), true, false) AS author#84, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 2, title), StringType), true, false) AS title#85, if (assertnotnull(input[0, org.apache.spark.sql.Row, true]).isNullAt) null else staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 3, incipit), StringType), true, false) AS incipit#86, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 4, to), BooleanType) AS to#87, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 5, get), BooleanType) AS get#88, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 6, alike), BooleanType) AS alike#89, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 7, with), BooleanType) AS with#90, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 8, man), BooleanType) AS man#91, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 9, in), BooleanType) AS in#92, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 10, anna), BooleanType) AS anna#93, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 11, was), BooleanType) AS was#94, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 12, happy), BooleanType) AS happy#95, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 13, pig), BooleanType) AS pig#96, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 14, way), BooleanType) AS way#97, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 15, morning), BooleanType) AS morning#98, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 16, city), BooleanType) AS city#99, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 17, when), BooleanType) AS when#100, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 18, two), BooleanType) AS two#101, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 19, drunk), BooleanType) AS drunk#102, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 20, pop), BooleanType) AS pop#103, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 21, John), BooleanType) AS John#104, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 22, four), BooleanType) AS four#105, validateexternaltype(getexternalrowfield(assertnotnull(input[0, org.apache.spark.sql.Row, true]), 23, hence), BooleanType) AS hence#106, ... 20 more fields]
+- *(1) MapElements , obj#82: org.apache.spark.sql.Row
   +- *(1) DeserializeToObject createexternalrow(id#9, author#10.toString, title#11.toString, incipit#12.toString, StructField(id,IntegerType,false), StructField(author,StringType,true), StructField(title,StringType,true), StructField(incipit,StringType,true)), obj#81: org.apache.spark.sql.Row
      +- InMemoryTableScan [id#9, author#10, title#11, incipit#12]
            +- InMemoryRelation [id#9, author#10, title#11, incipit#12], true, 10000, StorageLevel(disk, memory, deserialized, 1 replicas)
                  +- LocalTableScan [id#9, author#10, title#11, incipit#12]

Whatever the root cause is, the conclusion is clear. As per Spark 2.3.0 (and probably previous versions) adding (dynamically) a congruous number of columns to a dataframe should be done via a map operation and not foldLeft for the reasons we’ve seen.

—————————————-

^{1. [I run the tests on a virtual box with three cores running Spark 2.3.0. Time in milliseconds reflects the underline infrastructure and I would expect different performance on a proper cluster. Nevertheless, that a foldLeft solution performs significantly worse that a map based solution should be an outcome independent of the underline hardware.]↩}