[SPARK-17853] Kafka OffsetOutOfRangeException on DStreams union from separate Kafka clusters with identical topic names.

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.0.0
• Fix Version/s: 2.0.2, 2.1.0
• Component/s: DStreams
• Assignee: Cody Koeninger
• Reporter: Marcin Kuthan
• Created: 10/Oct/16 12:30
• Updated: 12/Oct/16 07:39
• Resolved: 12/Oct/16 07:39

Description

During migration from Spark 1.6 to 2.0 I observed OffsetOutOfRangeException reported by Kafka client. In our scenario we create single DStream as a union of multiple DStreams. One DStream for one Kafka cluster (multi dc solution). Both Kafka clusters have the same topics and number of partitions.

After quick investigation, I found that class DirectKafkaInputDStream keeps offset state for topic and partitions, but it is not aware of different Kafka clusters.

For every topic, single DStream is created as a union from all configured Kafka clusters.

class KafkaDStreamSource(configs: Iterable[Map[String, String]]) {
def createSource(ssc: StreamingContext, topic: String): DStream[(String, Array[Byte])] = {
    val streams = configs.map { config =>
      val kafkaParams = config
      val kafkaTopics = Set(topic)

      KafkaUtils.
          createDirectStream[String, Array[Byte]](
        ssc,
        LocationStrategies.PreferConsistent,
        ConsumerStrategies.Subscribe[String, Array[Byte]](kafkaTopics, kafkaParams)
      ).map { record =>
        (record.key, record.value)
      }
    }

    ssc.union(streams.toSeq)
  }
}

At the end, offsets from one Kafka cluster overwrite offsets from second one. Fortunately OffsetOutOfRangeException was thrown because offsets in both Kafka clusters are significantly different.

Issue Links

• links to: [Github] Pull Request #15442 (koeninger)