Avro Data Source for Apache Spark

A library for reading and writing Avro data from Spark SQL.

Requirements

This documentation is for version 3.2.0 of this library, which supports Spark 2.0+. For documentation on earlier versions of this library, see the links below.

This library has different versions for Spark 1.2, 1.3, 1.4 through 1.6, and 2.0+:

Spark Version	Compatible version of Avro Data Source for Spark
1.2	0.2.0
1.3	1.0.0
1.4-1.6	2.0.1
2.0+	3.2.0 (this version)

Linking

This library is cross-published for Scala 2.11, so 2.11 users should replace 2.10 with 2.11 in the commands listed below.

You can link against this library in your program at the following coordinates:

Using SBT:

libraryDependencies += "com.databricks" %% "spark-avro" % "3.2.0"

Using Maven:

<dependency>
    <groupId>com.databricks</groupId>
    <artifactId>spark-avro_2.10</artifactId>
    <version>3.2.0</version>
</dependency>

With spark-shell or spark-submit

This library can also be added to Spark jobs launched through spark-shell or spark-submit by using the --packages command line option. For example, to include it when starting the spark shell:

$ bin/spark-shell --packages com.databricks:spark-avro_2.11:3.2.0

Unlike using --jars, using --packages ensures that this library and its dependencies will be added to the classpath. The --packages argument can also be used with bin/spark-submit.

Features

Avro Data Source for Spark supports reading and writing of Avro data from Spark SQL.

• Automatic schema conversion: It supports most conversions between Spark SQL and Avro records, making Avro a first-class citizen in Spark.
• Partitioning: This library allows developers to easily read and write partitioned data witout any extra configuration. Just pass the columns you want to partition on, just like you would for Parquet.
• Compression: You can specify the type of compression to use when writing Avro out to disk. The supported types are uncompressed, snappy, and deflate. You can also specify the deflate level.
• Specifying record names: You can specify the record name and namespace to use by passing a map of parameters with recordName and recordNamespace.

Supported types for Avro -> Spark SQL conversion

This library supports reading all Avro types. It uses the following mapping from Avro types to Spark SQL types:

Avro type	Spark SQL type
boolean	BooleanType
int	IntegerType
long	LongType
float	FloatType
double	DoubleType
bytes	BinaryType
string	StringType
record	StructType
enum	StringType
array	ArrayType
map	MapType
fixed	BinaryType
union	See below

In addition to the types listed above, it supports reading union types. The following three types are considered basic union types:

1. union(int, long) will be mapped to LongType.
2. union(float, double) will be mapped to DoubleType.
3. union(something, null), where something is any supported Avro type. This will be mapped to the same Spark SQL type as that of something, with nullable set to true.

All other union types are considered complex. They will be mapped to StructType where field names are member0, member1, etc., in accordance with members of the union. This is consistent with the behavior when converting between Avro and Parquet.

At the moment, it ignores docs, aliases and other properties present in the Avro file.

Supported types for Spark SQL -> Avro conversion

This library supports writing of all Spark SQL types into Avro. For most types, the mapping from Spark types to Avro types is straightforward (e.g. IntegerType gets converted to int); however, there are a few special cases which are listed below:

Spark SQL type	Avro type
ByteType	int
ShortType	int
DecimalType	string
BinaryType	bytes
TimestampType	long
StructType	record

Examples

The recommended way to read or write Avro data from Spark SQL is by using Spark's DataFrame APIs, which are available in Scala, Java, Python, and R.

These examples use an Avro file available for download here:

Scala API

// import needed for the .avro method to be added
import com.databricks.spark.avro._
import org.apache.spark.sql.SparkSession

val spark = SparkSession.builder().master("local").getOrCreate()

// The Avro records get converted to Spark types, filtered, and
// then written back out as Avro records
val df = spark.read.avro("src/test/resources/episodes.avro")
df.filter("doctor > 5").write.avro("/tmp/output")

Alternatively you can specify the format to use instead:

val spark = SparkSession.builder().master("local").getOrCreate()
val df = spark.read
    .format("com.databricks.spark.avro")
    .load("src/test/resources/episodes.avro")

df.filter("doctor > 5").write.format("com.databricks.spark.avro").save("/tmp/output")

You can specify a custom Avro schema:

import org.apache.avro.Schema
import org.apache.spark.sql.SparkSession

val schema = new Schema.Parser().parse(new File("user.avsc"))
val spark = SparkSession.builder().master("local").getOrCreate()
spark
  .read
  .format("com.databricks.spark.avro")
  .option("avroSchema", schema.toString)
  .load("src/test/resources/episodes.avro").show()

You can also specify Avro compression options:

import com.databricks.spark.avro._
import org.apache.spark.sql.SparkSession

val spark = SparkSession.builder().master("local").getOrCreate()

// configuration to use deflate compression
spark.conf.set("spark.sql.avro.compression.codec", "deflate")
spark.conf.set("spark.sql.avro.deflate.level", "5")

val df = spark.read.avro("src/test/resources/episodes.avro")

// writes out compressed Avro records
df.write.avro("/tmp/output")

You can write partitioned Avro records like this:

import com.databricks.spark.avro._
import org.apache.spark.sql.SparkSession

val spark = SparkSession.builder().master("local").getOrCreate()

val df = spark.createDataFrame(
  Seq(
    (2012, 8, "Batman", 9.8),
    (2012, 8, "Hero", 8.7),
    (2012, 7, "Robot", 5.5),
    (2011, 7, "Git", 2.0))
  ).toDF("year", "month", "title", "rating")

df.toDF.write.partitionBy("year", "month").avro("/tmp/output")

You can specify the record name and namespace like this:

import com.databricks.spark.avro._
import org.apache.spark.sql.SparkSession

val spark = SparkSession.builder().master("local").getOrCreate()
val df = spark.read.avro("src/test/resources/episodes.avro")

val name = "AvroTest"
val namespace = "com.databricks.spark.avro"
val parameters = Map("recordName" -> name, "recordNamespace" -> namespace)

df.write.options(parameters).avro("/tmp/output")

Java API

import org.apache.spark.sql.*;
import org.apache.spark.sql.functions;

SparkSession spark = SparkSession.builder().master("local").getOrCreate();

// Creates a DataFrame from a specified file
Dataset<Row> df = spark.read().format("com.databricks.spark.avro")
  .load("src/test/resources/episodes.avro");

// Saves the subset of the Avro records read in
df.filter(functions.expr("doctor > 5")).write()
  .format("com.databricks.spark.avro")
  .save("/tmp/output");

Python API

# Creates a DataFrame from a specified directory
df = spark.read.format("com.databricks.spark.avro").load("src/test/resources/episodes.avro")

#  Saves the subset of the Avro records read in
subset = df.where("doctor > 5")
subset.write.format("com.databricks.spark.avro").save("/tmp/output")

SQL API

Avro data can be queried in pure SQL by registering the data as a temporary table.

CREATE TEMPORARY TABLE episodes
USING com.databricks.spark.avro
OPTIONS (path "src/test/resources/episodes.avro")

Building From Source

This library is built with SBT, which is automatically downloaded by the included shell script. To build a JAR file simply run build/sbt package from the project root.

Testing

To run the tests, you should run build/sbt test. In case you are doing improvements that target speed, you can generate a sample Avro file and check how long it takes to read that Avro file using the following commands:

build/sbt "test:run-main com.databricks.spark.avro.AvroFileGenerator NUMBER_OF_RECORDS NUMBER_OF_FILES"

will create sample avro files in target/avroForBenchmark/. You can specify the number of records for each file, as well as the overall number of files.

build/sbt "test:run-main com.databricks.spark.avro.AvroReadBenchmark"

runs count() on the data inside target/avroForBenchmark/ and tells you how the operation took.

Similarly, you can do benchmarks on how long it takes to write DataFrame as Avro file with

build/sbt "test:run-main com.databricks.spark.avro.AvroWriteBenchmark NUMBER_OF_ROWS"

where NUMBER_OF_ROWS is an optional parameter that allows you to specify the number of rows in DataFrame that we will be writing.