Learning PySpark: Build data-intensive applications locally and deploy at scale using the combined powers of Python and Spark 2.0

Any Spark application spins off a single driver process (that can contain multiple jobs) on the master node that then directs executor processes (that contain multiple tasks) distributed to a number of worker nodes as noted in the following diagram:

The driver process determines the number and the composition of the task processes directed to the executor nodes based on the graph generated for the given job. Note, that any worker node can execute tasks from a number of different jobs.

A Spark job is associated with a chain of object dependencies organized in a direct acyclic graph (DAG) such as the following example generated from the Spark UI. Given this, Spark can optimize the scheduling (for example, determine the number of tasks and workers required) and execution of these tasks:

Apache Spark is built around a distributed collection of immutable Java Virtual Machine (JVM) objects called Resilient Distributed Datasets (RDDs for short). As we are working with Python, it is important to note that the Python data is stored within these JVM objects. More of this will be discussed in the subsequent chapters on RDDs and DataFrames. These objects allow any job to perform calculations very quickly. RDDs are calculated against, cached, and stored in-memory: a scheme that results in orders of magnitude faster computations compared to other traditional distributed frameworks like Apache Hadoop.

At the same time, RDDs expose some coarse-grained transformations (such as map(...), reduce(...), and filter(...) which we will cover in greater detail in Chapter 2, Resilient Distributed Datasets), keeping the flexibility and extensibility of the Hadoop platform to perform a wide variety of calculations. RDDs apply and log transformations to the data in parallel, resulting in both increased speed and fault-tolerance. By registering the transformations, RDDs provide data lineage - a form of an ancestry tree for each intermediate step in the form of a graph. This, in effect, guards the RDDs against data loss - if a partition of an RDD is lost it still has enough information to recreate that partition instead of simply depending on replication.

RDDs have two sets of parallel operations: transformations (which return pointers to new RDDs) and actions (which return values to the driver after running a computation); we will cover these in greater detail in later chapters.

RDD transformation operations are lazy in a sense that they do not compute their results immediately. The transformations are only computed when an action is executed and the results need to be returned to the driver. This delayed execution results in more fine-tuned queries: Queries that are optimized for performance. This optimization starts with Apache Spark's DAGScheduler – the stage oriented scheduler that transforms using stages as seen in the preceding screenshot. By having separate RDD transformations and actions, the DAGScheduler can perform optimizations in the query including being able to avoid shuffling, the data (the most resource intensive task).

For more information on the DAGScheduler and optimizations (specifically around narrow or wide dependencies), a great reference is the Narrow vs. Wide Transformations section in High Performance Spark in Chapter 5, Effective Transformations (https://smile.amazon.com/High-Performance-Spark-Practices-Optimizing/dp/1491943203).

DataFrames, like RDDs, are immutable collections of data distributed among the nodes in a cluster. However, unlike RDDs, in DataFrames data is organized into named columns.

DataFrames were designed to make large data sets processing even easier. They allow developers to formalize the structure of the data, allowing higher-level abstraction; in that sense DataFrames resemble tables from the relational database world. DataFrames provide a domain specific language API to manipulate the distributed data and make Spark accessible to a wider audience, beyond specialized data engineers.

One of the major benefits of DataFrames is that the Spark engine initially builds a logical execution plan and executes generated code based on a physical plan determined by a cost optimizer. Unlike RDDs that can be significantly slower on Python compared with Java or Scala, the introduction of DataFrames has brought performance parity across all the languages.

Introduced in Spark 1.6, the goal of Spark Datasets is to provide an API that allows users to easily express transformations on domain objects, while also providing the performance and benefits of the robust Spark SQL execution engine. Unfortunately, at the time of writing this book Datasets are only available in Scala or Java. When they are available in PySpark we will cover them in future editions.

Spark SQL is one of the most technically involved components of Apache Spark as it powers both SQL queries and the DataFrame API. At the core of Spark SQL is the Catalyst Optimizer. The optimizer is based on functional programming constructs and was designed with two purposes in mind: To ease the addition of new optimization techniques and features to Spark SQL and to allow external developers to extend the optimizer (for example, adding data source specific rules, support for new data types, and so on):

Tungsten is the codename for an umbrella project of Apache Spark's execution engine. The project focuses on improving the Spark algorithms so they use memory and CPU more efficiently, pushing the performance of modern hardware closer to its limits.

The efforts of this project focus, among others, on:

In the previous section, we stated out that Datasets (at the time of writing this book) are only available in Scala or Java. However, we are providing the following context to better understand the direction of Spark 2.0.

Datasets were introduced in 2015 as part of the Apache Spark 1.6 release. The goal for datasets was to provide a type-safe, programming interface. This allowed developers to work with semi-structured data (like JSON or key-value pairs) with compile time type safety (that is, production applications can be checked for errors before they run). Part of the reason why Python does not implement a Dataset API is because Python is not a type-safe language.

Just as important, the Datasets API contain high-level domain specific language operations such as sum(), avg(), join(), and group(). This latter trait means that you have the flexibility of traditional Spark RDDs but the code is also easier to express, read, and write. Similar to DataFrames, Datasets can take advantage of Spark's catalyst optimizer by exposing expressions and data fields to a query planner and making use of Tungsten's fast in-memory encoding.

The history of the Spark APIs is denoted in the following diagram noting the progression from RDD to DataFrame to Dataset:

Source: From Webinar Apache Spark 1.5: What is the difference between a DataFrame and a RDD? http://bit.ly/29JPJSA

The unification of the DataFrame and Dataset APIs has the potential of creating breaking changes to backwards compatibility. This was one of the main reasons Apache Spark 2.0 was a major release (as opposed to a 1.x minor release which would have minimized any breaking changes). As you can see from the following diagram, DataFrame and Dataset both belong to the new Dataset API introduced as part of Apache Spark 2.0:

Source: A Tale of Three Apache Spark APIs: RDDs, DataFrames, and Datasets http://bit.ly/2accSNA

As noted previously, the Dataset API provides a type-safe, object-oriented programming interface. Datasets can take advantage of the Catalyst optimizer by exposing expressions and data fields to the query planner and Project Tungsten's Fast In-memory encoding. But with DataFrame and Dataset now unified as part of Apache Spark 2.0, DataFrame is now an alias for the Dataset Untyped API. More specifically:

DataFrame = Dataset[Row]

In the past, you would potentially work with SparkConf, SparkContext, SQLContext, and HiveContext to execute your various Spark queries for configuration, Spark context, SQL context, and Hive context respectively. The SparkSession is essentially the combination of these contexts including StreamingContext.

For example, instead of writing:

df = sqlContext.read \
    .format('json').load('py/test/sql/people.json')

now you can write:

df = spark.read.format('json').load('py/test/sql/people.json')

or:

df = spark.read.json('py/test/sql/people.json')

The SparkSession is now the entry point for reading data, working with metadata, configuring the session, and managing the cluster resources.

The fundamental observation of the computer hardware landscape when the project started was that, while there were improvements in price per performance in RAM memory, disk, and (to an extent) network interfaces, the price per performance advancements for CPUs were not the same. Though hardware manufacturers could put more cores in each socket (i.e. improve performance through parallelization), there were no significant improvements in the actual core speed.

Project Tungsten was introduced in 2015 to make significant changes to the Spark engine with the focus on improving performance. The first phase of these improvements focused on the following facets:

The following diagram is the updated Catalyst engine to denote the inclusion of Datasets. As you see at the right of the diagram (right of the Cost Model), Code Generation is used against the selected physical plans to generate the underlying RDDs:

Source: Structuring Spark: DataFrames, Datasets, and Streaming http://bit.ly/2cJ508x

As part of Tungsten Phase 2, there is the push into whole-stage code generation. That is, the Spark engine will now generate the byte code at compile time for the entire Spark stage instead of just for specific jobs or tasks. The primary facets surrounding these improvements include:

For a more in-depth review of Project Tungsten, please refer to:

As quoted by Reynold Xin during Spark Summit East 2016:

This is the underlying foundation for building Structured Streaming. While streaming is powerful, one of the key issues is that streaming can be difficult to build and maintain. While companies such as Uber, Netflix, and Pinterest have Spark Streaming applications running in production, they also have dedicated teams to ensure the systems are highly available.

As implied previously, there are many things that can go wrong when operating Spark Streaming (and any streaming system for that matter) including (but not limited to) late events, partial outputs to the final data source, state recovery on failure, and/or distributed reads/writes:

Source: A Deep Dive into Structured Streaming http://bit.ly/2aHt1w0

Therefore, to simplify Spark Streaming, there is now a single API that addresses both batch and streaming within the Apache Spark 2.0 release. More succinctly, the high-level streaming API is now built on top of the Apache Spark SQL Engine. It runs the same queries as you would with Datasets/DataFrames providing you with all the performance and optimization benefits as well as benefits such as event time, windowing, sessions, sources, and sinks.

Altogether, Apache Spark 2.0 not only unified DataFrames and Datasets but also unified streaming, interactive, and batch queries. This opens a whole new set of use cases including the ability to aggregate data into a stream and then serving it using traditional JDBC/ODBC, to change queries at run time, and/or to build and apply ML models in for many scenario in a variety of latency use cases:

Source: Apache Spark Key Terms, Explained https://databricks.com/blog/2016/06/22/apache-spark-key-terms-explained.html.

Together, you can now build end-to-end continuous applications, in which you can issue the same queries to batch processing as to real-time data, perform ETL, generate reports, update or track specific data in the stream.