RDD, DataFrame or Dataset

总结： 
1、RDD是一个Java对象的集合。RDD的优点是更面向对象，代码更容易理解。但在需要在集群中传输数据时需要为每个对象保留数据及结构信息，这会导致数据的冗余，同时这会导致大量的GC。 
2、DataFrame是在1.3引入的，它包含数据与schema2部分信息，其中数据就是真正的数据，而不是一个java对象。它不容易理解，同时对java支持不好，还有一个缺点是非强类型，这会导致部分错误在运行时才会发现。优点是数据不需要加载到一个java对象，减少GC，大大优化了数据在集群间传播与本地序列化的效率。 
3、DataSet在1.6引入了预览版，在2.0才真正稳定。它试图整合RDD/DataFrame的优点。在2.0里对DataSet的定位是：（1）DataFrame只是一个type alias，真正实现都是DataSet。（2）对于Python和R这些非类型安全的语言，DataFrame仍是主要编程接口。

• Unifying DataFrames and Datasets in Scala/Java: Starting in Spark 2.0, DataFrame is just a type alias for Dataset of Row. Both the typed methods (e.g. map, filter, groupByKey) and the untyped methods (e.g. select, groupBy) are available on the Dataset class. Also, this new combined Dataset interface is the abstraction used for Structured Streaming. Since compile-time type-safety in Python and R is not a language feature, the concept of Dataset does not apply to these languages’ APIs. Instead, DataFrame remains the primary programing abstraction, which is analogous to the single-node data frame notion in these languages. Get a peek from a Dataset API notebook.

• DataFrame-based Machine Learning API emerges as the primary ML API: With Spark 2.0, the spark.ml package, with its “pipeline” APIs, will emerge as the primary machine learning API. While the original spark.mllib package is preserved, future development will focus on the DataFrame-based API.

There Are Now 3 Apache Spark APIs. Here’s How to Choose the Right One

See Apache Spark 2.0 API Improvements: RDD, DataFrame, DataSet and SQL here.

Apache Spark is evolving at a rapid pace, including changes and additions to core APIs. One of the most disruptive areas of change is around the representation of data sets. Spark 1.0 used the RDD API but in the past twelve months, two new alternative and incompatible APIs have been introduced. Spark 1.3 introduced the radically different DataFrame API and the recently released Spark 1.6 release introduces a preview of the new Dataset API.

Many existing Spark developers will be wondering whether to jump from RDDs directly to the Dataset API, or whether to first move to the DataFrame API. Newcomers to Spark will have to choose which API to start learning with.

This article provides an overview of each of these APIs, and outlines the strengths and weaknesses of each one. A companion github repository provides working examples that are a good starting point for experimentation with the approaches outlined in this article.

The RDD (Resilient Distributed Dataset) API has been in Spark since the 1.0 release. This interface and its Java equivalent, JavaRDD, will be familiar to any developers who have worked through the standard Spark tutorials. From a developer’s perspective, an RDD is simply a set of Java or Scala objects representing data.

The RDD API provides many transformation methods, such as map(), filter(), and reduce() for performing computations on the data. Each of these methods results in a new RDD representing the transformed data. However, these methods are just defining the operations to be performed and the transformations are not performed until an action method is called. Examples of action methods are collect() and saveAsObjectFile().

Example of RDD transformations and actions

Scala:

rdd.filter(_.age > 21)              // transformation
   .map(_.last)                     // transformation                     
   .saveAsObjectFile("under21.bin") // action

Java:

rdd.filter(p -> p.getAge() < 21)     // transformation 
   .map(p -> p.getLast())            // transformation
   .saveAsObjectFile("under21.bin"); // action

The main advantage of RDDs is that they are simple and well understood because they deal with concrete classes, providing a familiar object-oriented programming style with compile-time type-safety. For example, given an RDD containing instances of Person we can filter by age by referencing the age attribute of each Person object:

Example: Filter by attribute with RDD

Scala:

rdd.filter(_.age > 21)

Java:

rdd.filter(person -> person.getAge() > 21)

Because the code is referring to data attributes by name, it is not possible for the compiler to catch any errors. If attribute names are incorrect then the error will only detected at runtime, when the query plan is created.

Another downside with the DataFrame API is that it is very scala-centric and while it does support Java, the support is limited. For example, when creating a DataFrame from an existing RDD of Java objects, Spark’s Catalyst optimizer cannot infer the schema and assumes that any objects in the DataFrame implement the scala.Product interface. Scala case classes work out the box because they implement this interface.

Dataset API

The Dataset API, released as an API preview in Spark 1.6, aims to provide the best of both worlds; the familiar object-oriented programming style and compile-time type-safety of the RDD API but with the performance benefits of the Catalyst query optimizer. Datasets also use the same efficient off-heap storage mechanism as the DataFrame API.

When it comes to serializing data, the Dataset API has the concept of encoders which translate between JVM representations (objects) and Spark’s internal binary format. Spark has built-in encoders which are very advanced in that they generate byte code to interact with off-heap data and provide on-demand access to individual attributes without having to de-serialize an entire object. Spark does not yet provide an API for implementing custom encoders, but that is planned for a future release.

Additionally, the Dataset API is designed to work equally well with both Java and Scala. When working with Java objects, it is important that they are fully bean-compliant. In writing the examples to accompany this article, we ran into errors when trying to create a Dataset in Java from a list of Java objects that were not fully bean-compliant.

Example: Creating Dataset from a list of objects

Scala

val sc = new SparkContext(conf)
val sqlContext = new SQLContext(sc)
import sqlContext.implicits._
val sampleData: Seq[ScalaPerson] = ScalaData.sampleData()
val dataset = sqlContext.createDataset(sampleData)

Java

JavaSparkContext sc = new JavaSparkContext(sparkConf);
SQLContext sqlContext = new SQLContext(sc);
List data = JavaData.sampleData();
Dataset dataset = sqlContext.createDataset(data, Encoders.bean(JavaPerson.class));

Transformations with the Dataset API look very much like the RDD API and deal with the Person class rather than an abstraction of a row.

Example: Filter by attribute with Dataset

Scala

dataset.filter(_.age < 21);

Java

dataset.filter(person -> person.getAge() < 21);

Despite the similarity with RDD code, this code is building a query plan, rather than dealing with individual objects, and if age is the only attribute accessed, then the rest of the the object’s data will not be read from off-heap storage.

Conclusions

If you are developing primarily in Java then it is worth considering a move to Scala before adopting the DataFrame or Dataset APIs. Although there is an effort to support Java, Spark is written in Scala and the code often makes assumptions that make it hard (but not impossible) to deal with Java objects.

If you are developing in Scala and need your code to go into production with Spark 1.6.0 then the DataFrame API is clearly the most stable option available and currently offers the best performance.

However, the Dataset API preview looks very promising and provides a more natural way to code. Given the rapid evolution of Spark it is likely that this API will mature very quickly through 2016 and become the de-facto API for developing new applications.

See Apache Spark 2.0 API Improvements: RDD, DataFrame, DataSet and SQL here.