[Coursera] Big Data Analysis with Scala and Spark 课程笔记

Note:
This blog is intended to record what I learned in Big Data Analysis with Scala and Spark, rather than share the course materials. So there are two things that I won’t put up: my solution to assignments and topics which I’ve mastered already.
PLEASE NOTIFY ME if I broke Coursera Honor Code accidentally.

Week 1

Introduction, Logistics, What You’ll Learn

  • this course is about distributed data parallelism in Spark.
  • the contents of this course:
    • extending data parallel paradigm to the distributed case, using Spark.
    • Spark’s programming model.
    • distributing computation, and cluster topology in Spark.
    • how to improve performance; data locality, how to avoid recomputation and shuffles in Spark.
    • relational operations with DataFrames and Datasets.
  • recommended books and resouces:
    • Learning Spark (2015) (Holden Karau): covering the basics.
    • Spark in Action (2017) (Petar Zecevic): covering the basics, full of examples, first time learning.
    • High Performance Spark (Holden Karau): deeper, how Spark executes jobs, how to squeeze better performance.
    • Advanced Analytics with Spark (2015) (Sandy Ryza): mainly focusing on data science.
    • Mastering Apache Spark 2 (Jacek Laskowski).

Data-Parallel to Distributed Data-Parallel

  • shared memory data parallelism:
    • split the data.
    • workers/threads independently operate on the data shards in parallel.
    • combine when done (if necessary) .
  • distributed data parallelism:
    • split the data over several nodes.
    • nodes independently operate on the data shards in parallel.
    • combine when done ( if necessary).
  • distributed data parallalism has a new concern: network latency between workers.
  • Spark implements a distributed data parallel model called Resilient Distributed Datasets (RDDs).

Latency

  • distribution introduces important concerns beyond what we had to worry about when dealing with parallelism in the shared memory case:
    • partial failure: crash failures of a subset of the machines involved in a distributed computation.
    • latency: certain operations have a much higher latency than other operations due to network communication.
  • latency cannot be masked completely.
  • fault-tolerance in Hadoop/MapReduce comes at a cost.
    • between each map and reduce step, in order to recover from potential failures, Hadoop/MapReduce shuffles its data and write intermediate data to disk.
  • while spark achieves fault tolerance using ideas from functional programming.
    • keep all data immutable and in-memory.
    • all operations on data are just functional transformations, like regular Scala collections.
    • fault tolerance is achieved by replaying functional transformations over original dataset.

RDDs, Spark’s Distributed Collection

  • RDDs seem a lot like immutable sequential or parallel Scala collections.
    • combinators on RDDs: map, flatMap, filter, reduce, fold, aggregate, …
  • difference between two signatures of aggregate methods:

    1
    2
    3
    4
    5
    6
    7
    8
    aggregate[B](z: => B)(seqop: (B, A)=> B, combop: (B, B) => B): B // Scala, by-name
    aggregate[B](z: B)(seqop: (B, A)=> B, combop: (B, B) => B): B // Spark RDD, by-value
    /**
    * I guess aggregate of Spark RDD does not use by-name evaluation because
    * the computation actually occurs on every worker node, which means the
    * initial value (by-name parameter) would be evaluated same times as
    * the number of worker nodes. It does cause performance loss, I guess.
    */
  • RDDs can be created in two ways:

    • transforming an existing RDD by high-order functions.
    • from a SparkContext (or SparkSession) object, which defines a handful of methods which can be used to create and populate a new RDD:
      1
      2
      parallelize // convert a local Scala collection to an RDD .
      textFile // read a text file from HDFS or a local file system and return an RDD of String

RDDs: Transformation and Actions

  • for Scala collection:
    • transformers: return new collections as results. map, filter, flatMap, groupBy.
    • accessors: return singles values. reduce, fold, aggregate.
  • similarly, Spark defines transformations and actions on RDDs:
    • transformations: return new RDDs. It’s lazy that its result is not immediately computed.
    • actions: compute a result based on an RDD, and either returned or saved to an external storage system. It’s eager (instead of lazy).
  • laziness/eagerness is how we can limit network communication using the programming model.
  • common transformations (lazy):
    • map: map[B](f: A => B): RDD[B].
    • flatMap: flatMap[B](f: A => TraversableOnce[B]).
    • filter: filter(pred: A => Boolean).
    • distinct: distinct(): RDD[B].
  • common actions (eager):
    • collect: collect(): Array[T], return all elements from RDD.
    • count: count(): Long, return the number of elements in the RDD.
    • take: take(n: Int): Array[T], return the first n elements of the RDD.
    • reduce: reduce(op: (A, A) => A).
    • foreach: foreach(f: T => Unit).
  • Spark computes RDDs the first time they are used in an action, which helps when processing large amounts of data.
  • Spark leverages this by analyzing and optimizing the chain of operations before executing it.

    1
    2
    3
    4
    5
    6
    val firstlogsWithErrors = lastYearslogs.filter(_.contains("ERROR")) .take(10)
    /**
    * The execution of filter is deferred until the take action is applied.
    * As soon as 10 elements of the filtered RDD have been computed,
    * firstLogsWithErrors is done.
    */
  • transformations on two RDDs:

    • union: union(other: RDD[T]): RDD[T].
    • intersection: intersection(other: RDD[T]): RDD[T].
    • subtract: subtract(other: RDD[T]): RDD[T].
    • cartesian: cartesian[U](other: RDD[U]): RDD[(T, U)].
  • other useful actions
    • takeSample: takeSample(withRepl: Boolean, num: Int): Array[T], return an array of a random sample of num elements of the dataset, with or without replacement.
    • takeOrdered: takeOrdered(num: Int)(implicit ord: Ordering[T]): Array[T], return the first n elements of the RDDs using either their natural order or a custom comparator.
    • saveAsTextFile: saveAsTextFile(path: String): Unit, write the elements of the dataset as a text file in the local filesystem or HDFS.
    • saveAsSequenceFile: saveAsSequenceFile(path: String): Unit, write the elements of the dataset as a Hadoop SequenceFile in the local filesystem or HDFS.

Evaluation in Spark: Unlike Scala Collections!

  • by default, RDDs are re-computed every time you run an action on them, which can be expensive if the dataset is used more than once.
  • luckily, Spark allows us to control what is cached in memory.

    • to tell Spark to cache an RDD in memory, simply call persist() or cache() on it.
      1
      2
      3
      4
      5
      val lastYearsLogs: RDD[String] = ...
      // logsWithErrors is cached in memory.
      val logsWithErrors = lastYearsLogs.filter(_.contains("error")).persist()
      val firstLogsWithErrors = logsWithErrors.take(10)
      val numErrors = logsWithErrors.count() // faster than no-caching
  • there are many ways to configure how your data is persisted.

    • in memory as regular Java objects.
    • on disk as regular Java objects.
    • in memory as serialized Java objects (more compact).
    • on disk as serialized regular Java objects (more compact).
    • both in memory and disk (spills over to disk to avoid re-computation).
  • one of the most common performance bottlenecks of newcomers to Spark arises from unknownly re-evaluating several transformations when caching coud be used.

Cluster Topology Matters!

  • a Spark application is a set of processes running on a cluster.
  • all these processes are coordinated by the driver program.
  • the driver is:
    • the process where the main() method of your program runs.
    • the process running the code that creates a SparkContext, creates RDDs, and stages up or sends off transformations and actions.
  • these processes that run computations and store data for your application are executors.
  • executors:
    • run the tasks that represent the application.
    • return computed results to the driver.
    • provide in-memory storage for cached RDDs.
  • execution of a Spark program:
    • the driver program runs the Spark application, which creates a SparkContext upon start-up.
    • the SparkContext connects to a cluster manager (e.g., Mesos/YARN) which allocates resources.
    • spark acquires executors on nodes in the cluster, which are processes that run computations and store data for your application.
    • next, driver program sends your application code to the executors.
    • finally, SparkContext sends tasks for the executors to run.
  • programmers should know where their code is running to avoid unexpected output or side-effects.

Week 2

Reduction Operations

  • reduction operations: walk through a collection and combine neighboring elements of the collection together to produce a single combined result.
  • properties of aggregate:
    • parallelizable. (like ‘fold’)
    • possible to change the return type. (like ‘foldLeft’)
  • Spark does not support serial foldLeft or foldRight because doing things serially across a cluster requires lots of synchronization, which is actually difficult.
  • as you will realize after experimenting with Spark a bit, much of the time when working with large-scale data, your goal is to project down from larger/more complex data types.

Pair RDDs

  • most common in world of big data processing: operating on data in the form of key-value pairs.
  • in Spark, distributed key-value pairs are PairRDD.

Transformations and Actions on Pair RDDs

  • transformations:
    • groupByKey: def groupByKey(): RDD[(K, Iterable[V])].
    • reduceByKey: def reduceByKey(func: (V, V) => V): RDD[(K, V)], can be thought of as a combination of groupByKey and reduce-ing on all the values per key. It’s more efficient though, than using each separately.
    • mapValues: def mapValues[U] (f: V => U) : RDD [(K, U)].
    • keys: def keys: RDD[K], return an RDD with the keys of each tuple.
    • join.
    • leftOuterJoin/rightOuterJoin.
  • actions:
    • countByKey: def countByKey(): Map[K, Long].

Joins

  • two kinds of joins:
    • inner joins (join).
    • outer joins (leftOuterJoin, rightOuterJoin).
  • the key difference between the two is what happens to the keys when both RDDs don’t contain the same key.
  • inner joins return a new RDD containing combined pairs whose keys are present in both input RDDs.
    • def join[W](other: RDD[(K, W)]): RDD[(K, (V, W))].
  • outer joins return a new RDD containing combined pairs whose keys don’t have to be present in both input RDDs.
    • def leftOuterJoin[W](other: RDD[(K, W)]): RDD[(K, (V, Option[W]))]
    • def rightOuterJoin[W](other: RDD[(K, W)]): RDD[(K, (Option[V], W))]

Week 3

Shuffling: What it is and why it’s important

  • shuffling: move data from one node to another to be grouped with its key.
    • shuffles can be an enormous hit due to network latency.
  • groupByKey results in one key-value pair per key. And this single key-value pair cannot span across multiple worker nodes.
  • conceptually, reduceByKey can be thought of as a combination of first doing groupByKey and then reduce-ing on all the values grouped per key.
    • but it’s more efficient though, than using each separately.
    • it is because an pre-reduceByKey can be done in each executor to reduce the amount of data sent over the network.
    • this can result in non-trival gains in performance!

Partitioning

  • partitioning: how Spark know which key to put on which machine. The data within an RDD is split into several partitions.
  • properties of partitions:
    • partitions never span multiple machines. i.e., tuples in the same partitions are guaranteed to be on the same machine.
    • each machine in the cluster contains one or more partitions.
    • the number of partitions to use is configurable. By default, it equals to the total number of cores on all executor nodes.
  • two kinds of partitioning available in Spark:
    • hash partitioning: attempts to spread data evenly across partitions based on the key.
    • range partitioning: may be more efficient when Pair RDDs contain keys that have an ordering defined (Int, String).
  • customizing a partitioning is only possible on Pair RDDs.
  • using a range partitioner, keys are partitioned according to:
    • an ordering of keys.
    • a set of sorted ranges of keys.
  • there are two ways to create RDDs with specific partitionings:
    • call partitionBy on an RDD, providing an explicit Partitioner.
    • using transformations that return RDDs with specific partitioners.

call partitionBy on an RDD:

1
2
3
val pairs = purchasesRdd.map(p => (p.customerld, p.price))
val tunedPartitioner = new RangePartitioner(8, pairs)
val partitioned = pairs.partitionBy(tunedPartitioner).persist()
  • creating a RangePartitioner requires:
    • specifying the desired number of partitions.
    • providing a Pair RDD with ordered keys. This RDD is sampled to create a suitable set of sorted ranges.
  • NOTE: the result of partitionBy should always be persisted. Otherwise, the partitioning is repeatedly applyed (involving shuffling) each time the partitioned RDD is used.

partitioning data using transformations:

  • partitioner from parent RDD:
    • Pair RDDs that are the result of a transformation on a partitioned Pair RDD typically is configured to use the hash partitioner that was used to construct it.
  • automatically-set partitioners:
    • some operations on RDDs automatically result in an RDD with a known partitioner - for when it makes sense.
    • by default, when using sortByKey, a RangePartitioner is used. Further, the default partitioner when using groupByKey, is a HashPartitioner.
  • operations on Pair RDDs that hold to (and propagate) a partitioner:
    • cogroup.
    • groupWith.
    • join.
    • leftOuterJoin.
    • rightOuterJoin.
    • groupByKey.
    • reduceByKey.
    • foldByKey.
    • combineByKey.
    • partitionBy.
    • sort.
    • mapValues (if parent has a partitioner).
    • flatMapValues (if parent has a partitioner).
    • filter (if parent has a partitioner).
  • all other operations will produce a result without a partitioner.
  • why map or flatMap do not preserve partitioner in their result RDDs?
    • because it’s possible for map or flatMap to change the key.
    • the previous partitioner would no longer make sense when key is changed.
    • hence mapValues enables us to still do map transformations without changing the keys, thereby preserving the partitioner.

Optimizing with Partitioners

  • using range partitioners we can optimize our earlier use of reduceByKey so that it does not involve any shuffling over the network at all!
  • grouping (groupByKey) is done using a hash partitioner with default parameters.
  • rule of thumb: a shuffle can occur when the resulting RDD depends on other elements from the same RDD or another RDD.
  • sometimes one can be clever and avoid much or all network communication while still using an operation like join via smart partitioning
  • we can figure out whether a shuffle has been planned/executed via:

    • the return type of certain transformations (ShuffledRDD).
    • using function toDebugString to see its execution plan.
      1
      2
      partitioned.reduceByKey((v1, v2) => (v1 ._1 + v2._1, v1 ._2 + v2._2))
      .toDebugString
  • operations that might cause a shuffle:

    • cogroup.
    • groupWith.
    • join.
    • leftOuterJoin.
    • rightOuterJoin.
    • groupByKey.
    • reduceByKey.
    • combineByKey.
    • distinct.
    • intersection.
    • repartition.
    • coalesce.
  • there are a few ways to use operations that might cause a shuffle and to still avoid much or all network shuffling.
    • reduceByKey running on a pre-partitioned ROD will cause the values to be computed locally, requiring only the final reduced value has to be sent from the worker to the driver.
    • join called on two RDDs that are pre-partitioned with the same partitioner and cached on the same machine will cause the join to be computed locally, with no shuffling across the network.

Wide vs Narrow Dependencies

  • computations on RDDs are represented as a lineage graph: a Directed Acyclic Graph (DAG) representing the computations done on the RDD.
  • RDDs are made up of 2 important parts (but are made up of 4 parts in total):
    • Partitions: atomic pieces of the dataset; one or many per compute node.
    • Dependencies. models relationship between this RDD and its partitions with the RDD(s) it was derived from.
    • A function for computing the dataset based on its parent RDDs.
    • Metadata about its partitioning scheme and data placement.
  • in fact, RDD dependencies encode when data must move across the network.
  • transformations can have two kinds of dependencies:
    • narrow dependencies: each partition of the parent RDD is used by at most one partition of the child RDD.
    • wide dependencies: each partition of the parent RDD may be depended on by multiple child partitions.
  • narrow dependencies mean the transformation can be:
    • fast.
    • no shuffle necessary.
    • optimizations like pipelining possible.
  • while wide dependencies are:
    • slow.
    • require all or some data to be shuffled over the network.
  • transformations with narrow dependencies:
    • map.
    • mapValues.
    • flatMap.
    • filter.
    • mapPartitions.
    • mapPartitionsWithIndex.
  • transformations with wide dependencies:
    • cogroup.
    • groupWith.
    • join.
    • leftOuterJoin.
    • rightOuterJoin.
    • groupByKey.
    • reduceByKey.
    • combineByKey.
    • distinct.
    • intersection.
    • repartition.
    • coalesce.
  • we could use dependencies method on RDD to return a sequence of denpendency objects, which are actually the dependencies used by Spark’s scheduler to know how this RDD depends on other RDDs.

    1
    2
    3
    val pairs = wordsRdd.map(c => (c, 1))
    .groupByKey()
    .dependencies
  • narrow dependency objects:

    • OneToOneDependency.
    • PruneDependency.
    • RangeDependency.
  • wide dependency objects:
    • ShuffleDependency.
  • toDebugString prints out a visualization of the RDD’s lineage, and other information pertinent to scheduling.
  • lineage graphs are the key to fault tolerance in Spark.
    • along with keeping track of dependency information between partitions as well, this allows us to:
    • recover from failures by recomputing lost partitions from lineage graphs.
  • recomputing missing partitions fast for narrow dependencies, but slow for wide dependencies.

Week 4

Structured vs Unstructured Data

  • all data isn’t equal, structurally. It falls on a spectrum from unstructured to structured.
    • unstructured: image, log files.
    • semi-structured: json, xml.
    • structured: database tables.
  • Spark + regular RDDs does not know anything about the schema of the data it’s dealing with, which makes it difficult to optimize aggressively.
  • Spark SQL makes it possible to do optimization work for users.

Spark SQL

  • Spark SQL supports:
    • seamlessly intermixing SQL queries with Scala.
    • all of optimizations we’re used in the databases community on Spark jobs.
  • three main goals of Spark SQL:
    • support relational processing both within Spark programs (on RDDs) and on external data sources with a friendly API.
    • high performance, achieved by using techniques from research in databases.
    • easily support new data sources such as semi-structured data and external databases.
  • DataFrames are, conceptually, RDDs full of records with a known schema.
  • DataFrames are untyped, which means Scala compiler won’t check the types in the schema.
  • transformations on DataFrames are also untyped.
  • DataFrames can be created in two ways:
    • from an existing RDD: either with schema inference, or with an explicit schema.
    • reading a specific data source from file: common structured or semi-structured formats such as JSON.
      1
      2
      3
      4
      5
      6
      val tupleRDD = ... // Assume RDD[(Int, String String, String)]
      val tupleDF = tupleRDD.toDF("id", "name", "city", "country") // column names
      // or
      case class Person(id: Int, name: String, city: String)
      val peopleRDD = ... // Assume RDD[Person]
      val peopleDF = peopleRDD.toDF

DataFrames (1)

  • complex Spark SQL data types:
Scala Type SQL Type
Array[T] ArrayType(elementType, containsNull)
Map[K, V] MapType(keyType, valueType, valueContainsNull)
case class StructType(List[StructFields])
  • we can select and work with columns in these ways:
    • using $-notation: df.filter($"age" > 18).
    • referring to the DataFrame: df.filter(df("age") > 18).
    • using SQL query string: df.filter("age > 18").
  • one of the most common tasks on tables is to (1) group data by a certain attribute, and then (2) do some kind of aggregation on it like a count.
  • for grouping & aggregating, Spark SQL provides:
    • a groupBy function which returns a RelationalGroupedDataset.
    • which has several standard aggregation functions defined on it like count, sum, max, min, and avg.

DataFrames (2)

  • dropping records with unwanted values:
    • drop(): drops rows that contain null or NaN values in any column and returns a new DF.
    • drop("all"): drops rows that contain null or NaN values in all columns and returns a new DF.
    • drop(Array("id", "name")): drops rows that contain null or NaN values in the specified columns and returns a new DF.
  • replacing unwanted values:
    • fill(0): replaces all occurrences of null or NaN in numeric columns with specified value and returns a new DF.
    • fill(Map("minBalance" -> 0)): replaces all occurrences of null or NaN in specified column with specified value and returns a new DF.
    • fill(Array("id"), Map(123 -> 1234)): replaces specified value (123) in specified column (“id”) with specified replacement value (1234) and returns a new DF.
  • common actions on DataFrames:
    • collect(): Array[Row]: returns an array that contains all of rows in this DataFrame.
    • count(): Long: returns the number of rows in the DataFrame.
    • first(): Row / head(): Row: return the first row in the DataFrame.
    • show(): Unit: displays the top 20 rows of the DataFrame in tabular form.
    • take(n: Int): Array[Row]: returns the first n rows in the DataFrame.
  • several types of joins are available: inner, outer, left_outer, right_outer, leftsemi.

    1
    2
    df1.join(df2, $"df1.id" === $"df2.id")
    df1.join(df2, $"df1.id" === $"df2.id", "right_outer")
  • Spark SQL comes with two specialized backend components:

    • Catalyst: query optimizer.
    • Tungsten: off-heap serializer.
  • Catalyst compiles Spark SQL programs down to an RDD.
    • reordering operations: laziness and structure gives us the ability to anaylyze and rearrange the DAG of computation.
    • reduce the amount of data we must read: skip reading in, serializing, and sending around parts of the data set that aren’t needed for computation.
    • pruning unneeded partitioning: analyze DataFrame and filter operations to figure out and skip partitions that are unneeded in computation.
  • Tungsten can provides:
    • highly-specialized data encoders: take schema information and tightly pack serialized data into memory.
    • column-based: when storing data, group data by column instead of row for faster lookups of data associated with specific attributes/columns.
    • off-heap (free from garbage collection overhead): regions of memory off heap, manually managed by Tungsten, so as to avoid GC overhead and pauses.
  • limitations of DataFrames:
    • untyped: Scala compiler does not check whether a column exists.
    • limited data types: difficult to ensure that a Tungsten encoder exist for your data type.
    • requires semi-structured/structured data: better to use regular RDDs for unstructured data.

Datasets

  • DataFrames are actually Datasets.

    1
    type DataFrame = Dataset[Row]
  • what is Dataset?

    • Datasets can be thought of as typed distributed collections of data.
    • Dataset API unifies the DataFrame and RDD APIs. Mix and match!
    • Datasets require structured/semi-structured data. Schemas and Encoders core part of Datasets.
  • we can think of Datasets as a compromise between RDDs and DataFrames:
    • more type information than DF.
    • more optimizations than RDDs.
  • the Dataset API includes both untyped and typed transformations:
    • untyped transformations: the transformations we learned on DataFrames.
    • typed transformations: typed variants of many DataFrame transformations + additional transformations such as RDD-like higher-order functions map, flatMap, etc.
  • common (typed) transformations on Datasets:
    • map: map[U](f: T => U): Dataset[U].
    • flatMap: flatMap[U](f: T => TraversableOnce[U]): Dataset[U].
    • filter: filter(pred: T => Boolean): Dataset[T].
    • distinct: distinct(): Dataset[T].
    • groupByKey: groupByKey[K](f: T => K): KeyValueGroupedDataset[K, T].
    • coalesce: coalesce(numPartitions: Int): Dataset[T].
    • repartition: repartition(numPartitions: Int): Dataset[T].
  • like on DataFrames, Datasets have a special set of aggregation operations meant to be used after a call to groupByKey on a Dataset .
    • calling groupByKey on a Dataset returns a KeyValueGroupedDataset.
    • KeyValueGroupedDataset contains a number of aggregation operations which return Datasets.
  • some KeyValueGroupedDataset aggregation operations:
    • reduceGroups: reduceGroups(f: (V, V) => V): Dataset[(K, V)].
    • agg: agg[U](col: TypedColumn[V, U]): Dataset[(K, U)].
    • mapGroups: mapGroups[U](f: (K, Iterator[V]) => U): Dataset[U].
    • flatMapGroups: flatMapGroups[U](f: (K, Iterator[V]) => TraversableOnce[U]): Dataset[U].
  • Aggregator: A class that helps you generically aggregate data. Kind of like the aggregate method we saw on RDDs.

    1
    2
    3
    4
    5
    6
    val myAgg = new Aggregator[IN, BUF, OUT] {
    def zero: BUF = ... // initial value
    def reduce(b: BUF, a: IN) : BUF = ... // add an element to the running total
    def merge(b1: BUF, b2: BUF): BUF = ... // merge the intermediate values
    def finish(b: BUF): OUT = ... // return the final result
    }.toColumn
  • Encoders are what convert your data between JVM objects and Spark SQL’s specialized internal (tabular) representation. They’re required by all Datasets!

  • two ways to introduce encoders:
    • automatically via implicits from a SparkSession: import spark.implicits._.
    • explicitly via org.apache.spark.sql.Encoder which contains a large selection of methods for creating Encoders from Scala primitive types and Proucts.
  • when to use DataFrames vs DataSets vs RDDs?
  • use Datasets when:
    • you have structured/ semi-structured data.
    • you want typesafety.
    • you need to work with functional APls.
    • you need good performance, but it doesn’t have to be the best.
  • use DataFrames when:
    • you have structured/semi-structured data.
    • you want the best possible performance, automatically optimized for you.
  • use RDDs when:
    • you have unstructured data.
    • you need to fine-tune and manage low-level details of RDD computations.
    • you have complex data types that cannot be serialized with Encoders.
  • limitations of Datasets:
    • Catalyst cannot optimize all operations.
    • all other limitations of DataFrames except untyped: limited data types and requiring semi-structured/structured data.
  • when using Datasets with higher-order functions like map, you miss out on many Catalyst optimizations .
  • when using Datasets with relational operations like select, you get all of Catalyst’s optimizations .
  • though not all operations on Datasets benefit from Catalyst’s optimizations, Tungsten is still always running under the hood of Datasets, storing and organizing data in a highly optimized way, which can result in large speedups over RDDs.