Analyze data with Spark

• 5 minutes

One of the benefits of using Spark is that you can write and run code in various programming languages, enabling you to use the programming skills you already have and to use the most appropriate language for a given task. The default language in a new Azure Synapse Analytics Spark notebook is PySpark - a Spark-optimized version of Python, which is commonly used by data scientists and analysts due to its strong support for data manipulation and visualization. Additionally, you can use languages such as Scala (a Java-derived language that can be used interactively) and SQL (a variant of the commonly used SQL language included in the Spark SQL library to work with relational data structures). Software engineers can also create compiled solutions that run on Spark using frameworks such as Java and Microsoft .NET.

Exploring data with dataframes

Natively, Spark uses a data structure called a resilient distributed dataset (RDD); but while you can write code that works directly with RDDs, the most commonly used data structure for working with structured data in Spark is the dataframe, which is provided as part of the Spark SQL library. Dataframes in Spark are similar to those in the ubiquitous Pandas Python library, but optimized to work in Spark's distributed processing environment.

Loading data into a dataframe

Let's explore a hypothetical example to see how you can use a dataframe to work with data. Suppose you have the following data in a comma-delimited text file named products.csv in the primary storage account for an Azure Synapse Analytics workspace:

ProductID,ProductName,Category,ListPrice
771,"Mountain-100 Silver, 38",Mountain Bikes,3399.9900
772,"Mountain-100 Silver, 42",Mountain Bikes,3399.9900
773,"Mountain-100 Silver, 44",Mountain Bikes,3399.9900
...

In a Spark notebook, you could use the following PySpark code to load the data into a dataframe and display the first 10 rows:

%%pyspark
df = spark.read.load('abfss://container@store.dfs.core.windows.net/products.csv',
    format='csv',
    header=True
)
display(df.limit(10))

The %%pyspark line at the beginning is called a magic, and tells Spark that the language used in this cell is PySpark. You can select the language you want to use as a default in the toolbar of the Notebook interface, and then use a magic to override that choice for a specific cell. For example, here's the equivalent Scala code for the products data example:

%%spark
val df = spark.read.format("csv").option("header", "true").load("abfss://container@store.dfs.core.windows.net/products.csv")
display(df.limit(10))

The magic %%spark is used to specify Scala.

Both of these code samples would produce output like this:

ProductID	ProductName	Category	ListPrice
771	Mountain-100 Silver, 38	Mountain Bikes	3399.9900
772	Mountain-100 Silver, 42	Mountain Bikes	3399.9900
773	Mountain-100 Silver, 44	Mountain Bikes	3399.9900
...	...	...	...

Specifying a dataframe schema

In the previous example, the first row of the CSV file contained the column names, and Spark was able to infer the data type of each column from the data it contains. You can also specify an explicit schema for the data, which is useful when the column names aren't included in the data file, like this CSV example:

771,"Mountain-100 Silver, 38",Mountain Bikes,3399.9900
772,"Mountain-100 Silver, 42",Mountain Bikes,3399.9900
773,"Mountain-100 Silver, 44",Mountain Bikes,3399.9900
...

The following PySpark example shows how to specify a schema for the dataframe to be loaded from a file named product-data.csv in this format:

from pyspark.sql.types import *
from pyspark.sql.functions import *

productSchema = StructType([
    StructField("ProductID", IntegerType()),
    StructField("ProductName", StringType()),
    StructField("Category", StringType()),
    StructField("ListPrice", FloatType())
    ])

df = spark.read.load('abfss://container@store.dfs.core.windows.net/product-data.csv',
    format='csv',
    schema=productSchema,
    header=False)
display(df.limit(10))

The results would once again be similar to:

ProductID	ProductName	Category	ListPrice
771	Mountain-100 Silver, 38	Mountain Bikes	3399.9900
772	Mountain-100 Silver, 42	Mountain Bikes	3399.9900
773	Mountain-100 Silver, 44	Mountain Bikes	3399.9900
...	...	...	...

Filtering and grouping dataframes

You can use the methods of the Dataframe class to filter, sort, group, and otherwise manipulate the data it contains. For example, the following code example uses the select method to retrieve the ProductName and ListPrice columns from the df dataframe containing product data in the previous example:

pricelist_df = df.select("ProductID", "ListPrice")

The results from this code example would look something like this:

ProductID	ListPrice
771	3399.9900
772	3399.9900
773	3399.9900
...	...

In common with most data manipulation methods, select returns a new dataframe object.

You can "chain" methods together to perform a series of manipulations that results in a transformed dataframe. For example, this example code chains the select and where methods to create a new dataframe containing the ProductName and ListPrice columns for products with a category of Mountain Bikes or Road Bikes:

bikes_df = df.select("ProductName", "ListPrice").where((df["Category"]=="Mountain Bikes") | (df["Category"]=="Road Bikes"))
display(bikes_df)

The results from this code example would look something like this:

ProductName	ListPrice
Mountain-100 Silver, 38	3399.9900
Road-750 Black, 52	539.9900
...	...

To group and aggregate data, you can use the groupBy method and aggregate functions. For example, the following PySpark code counts the number of products for each category:

counts_df = df.select("ProductID", "Category").groupBy("Category").count()
display(counts_df)

The results from this code example would look something like this:

Category	count
Headsets	3
Wheels	14
Mountain Bikes	32
...	...

Using SQL expressions in Spark

The Dataframe API is part of a Spark library named Spark SQL, which enables data analysts to use SQL expressions to query and manipulate data.

Creating database objects in the Spark catalog

The Spark catalog is a metastore for relational data objects such as views and tables. The Spark runtime can use the catalog to seamlessly integrate code written in any Spark-supported language with SQL expressions that may be more natural to some data analysts or developers.

One of the simplest ways to make data in a dataframe available for querying in the Spark catalog is to create a temporary view, as shown in the following code example:

df.createOrReplaceTempView("products")

A view is temporary, meaning that it's automatically deleted at the end of the current session. You can also create tables that are persisted in the catalog to define a database that can be queried using Spark SQL.

Using the Spark SQL API to query data

You can use the Spark SQL API in code written in any language to query data in the catalog. For example, the following PySpark code uses a SQL query to return data from the products view as a dataframe.

bikes_df = spark.sql("SELECT ProductID, ProductName, ListPrice \
                      FROM products \
                      WHERE Category IN ('Mountain Bikes', 'Road Bikes')")
display(bikes_df)

The results from the code example would look similar to the following table:

ProductID	ProductName	ListPrice
38	Mountain-100 Silver, 38	3399.9900
52	Road-750 Black, 52	539.9900
...	...	...

Using SQL code

The previous example demonstrated how to use the Spark SQL API to embed SQL expressions in Spark code. In a notebook, you can also use the %%sql magic to run SQL code that queries objects in the catalog, like this:

%%sql

SELECT Category, COUNT(ProductID) AS ProductCount
FROM products
GROUP BY Category
ORDER BY Category

The SQL code example returns a resultset that is automatically displayed in the notebook as a table, like the one below:

Category	ProductCount
Bib-Shorts	3
Bike Racks	1
Bike Stands	1
...	...