The issue here is related to the non-serializability of the avro.Schema class used in the Job. The exception is thrown when you try to reference the schema object from the code inside the map function.
For instance, if you try to do as follows, you will get the "Task not serializable" exception:
val schema = new Schema.Parser().parse(new File(jsonSchema))
...
rdd.map(t => {
// reference to the schema object declared outside
val record = new GenericData.Record(schema)
})
You can make everything to work by just creating a new instance of the schema inside the function block:
val schema = new Schema.Parser().parse(new File(jsonSchema))
// The schema above should not be used in closures, it's for other purposes
...
rdd.map(t => {
// create a new Schema object
val innserSchema = new Schema.Parser().parse(new File(jsonSchema))
val record = new GenericData.Record(innserSchema)
...
})
Since you would not like parsing the avro schema for every record you handle, a better solution will be to parse the schema at partition level. The following also works:
val schema = new Schema.Parser().parse(new File(jsonSchema))
// The schema above should not be used in closures, it's for other purposes
...
rdd.mapPartitions(tuples => {
// create a new Schema object
val innserSchema = new Schema.Parser().parse(new File(jsonSchema))
tuples.map(t => {
val record = new GenericData.Record(innserSchema)
...
// this closure will be bundled together with the outer one
// (no serialization issues)
})
})
The code above works as long as you provide a portable reference to the jsonSchema file, since the map function is going to be executed by multiple remote executors. It can be a reference to a file in HDFS or it can be packaged along with the application in the JAR (you will use the class-loader functions to get its contents in the latter case).
For those who are trying to use Avro with Spark, notice that there are still some unresolved compilation problems and you have to use the following import on Maven POM:
<dependency>
<groupId>org.apache.avro</groupId>
<artifactId>avro-mapred</artifactId>
<version>1.7.7</version>
<classifier>hadoop2</classifier>
<dependency>
Note the "hadoop2" classifier. You can track the issue at https://issues.apache.org/jira/browse/SPARK-3039.
