JavaScript Object Notation (JSON) is a data interchange format developed by the JavaScript ecosystem. It is a text-based format and has the same expressiveness such as, for instance, XML. The following example uses the SparkSession method called read.json to load the HDFS-based JSON data file named adult.json. This uses the so-called Apache Spark DataSource API to read and parse JSON files, but we will come back to that later.

val dframe = spark.read.json("hdfs:///data/spark/adult.json")

The result is a DataFrame.

Data can be saved in the JSON format using the DataSource API as well, as shown by the following example:

import spark.implicits._
val df = sc.parallelize(Array(1,2,3)).toDF
df.write.json("hdfs://localhost:9000/tmp/test.json")

So, the resulting data can be seen on HDFS; the Hadoop filesystem ls command shows you that the data resides in the target directory as a success file and eight part files. This is because even though small, the underlying RDD was set to have eight partitions, therefore those eight partitions have been written. This is shown in the following image:

What if we want to obtain a single file? This can be accomplished by repartition to a single partition:

val df1 =df.repartition(1)
df1.write.json("hdfs://localhost:9000/tmp/test_single_partition.json")

If we now have a look at the folder, it is a single file:

There are two important things to know. First, we still get the file wrapped in a subfolder, but this is not a problem as HDFS treats folders equal to files and as long as the containing files stick to the same format, there is no problem. So, if you refer to /tmp/test_single_partition.json, which is a folder, you can also use it similarly to a single file.

In addition, all files starting with _ are ignored. This brings us to the second point, the _SUCCESS file. This is a framework-independent way to tell users of that file that the job writing this file (or folder respectively) has been successfully completed. Using the Hadoop filesystem's cat command, it is possible to display the contents of the JSON data:

Processing Parquet data is very similar, as we will see next.