Accessing Avro data files from Spark SQL applications

Spark SQL supports loading and saving DataFrames from and to a variety of data sources. With the spark-avro library, you can process data encoded in the Avro format using Spark.

The spark-avro library supports most conversions between Spark SQL and Avro records, making Avro a first-class citizen in Spark. The library automatically performs the schema conversion. Spark SQL reads the data and converts it to Spark's internal representation; the Avro conversion is performed only during reading and writing data.

By default, when pointed at a directory, read methods silently skip any files that do not have the .avro extension. To include all files, set the avro.mapred.ignore.inputs.without.extension property to false. See Configuring Spark Applications.

Writing Compressed Data Files

To set the compression type used on write, configure the spark.sql.avro.compression.codec property:

sqlContext.setConf("spark.sql.avro.compression.codec","codec")

The supported codec values are uncompressed, snappy, and deflate. Specify the level to use with deflate compression in spark.sql.avro.deflate.level.

Accessing Partitioned Data Files

The spark-avro library supports writing and reading partitioned data. You pass the partition columns to the writer.

Specifying Record Name and Namespace

Specify the record name and namespace to use when writing to disk by passing recordName and recordNamespace as optional parameters.

Spark SQL

You can write SQL queries to query a set of Avro files. First, create a temporary table pointing to the directory containing the Avro files. Then query the temporary table:

sqlContext.sql("CREATE TEMPORARY TABLE table_name
  USING com.databricks.spark.avro OPTIONS (path "input_dir"))
df = sqlContext.sql("SELECT * FROM table_name")

Avro to Spark SQL Conversion

The spark-avro library supports conversion for all Avro data types:

boolean -> BooleanType
int -> IntegerType
long -> LongType
float -> FloatType
double -> DoubleType
bytes -> BinaryType
string -> StringType
record -> StructType
enum -> StringType
array -> ArrayType
map -> MapType
fixed -> BinaryType

The spark-avro library supports the following union types:

union(int, long) -> LongType
union(float, double) -> DoubleType
union(any, null) -> any

The library does not support complex union types.

All doc, aliases, and other fields are stripped when they are loaded into Spark.

Spark SQL to Avro Conversion

Every Spark SQL type is supported:

BooleanType -> boolean
IntegerType -> int
LongType -> long
FloatType -> float
DoubleType -> double
BinaryType -> bytes
StringType -> string
StructType -> record
ArrayType -> array
MapType -> map
ByteType -> int
ShortType -> int
DecimalType -> string
BinaryType -> bytes
TimestampType -> long

Limitations

Because Spark is converting data types, keep the following in mind:

Enumerated types are erased - Avro enumerated types become strings when they are read into Spark, because Spark does not support enumerated types.
Unions on output - Spark writes everything as unions of the given type along with a null option.
Avro schema changes - Spark reads everything into an internal representation. Even if you just read and then write the data, the schema for the output is different.
Spark schema reordering - Spark reorders the elements in its schema when writing them to disk so that the elements being partitioned on are the last elements.