diff --git a/src/main/scala/io/epiphanous/flinkrunner/serde/ConfluentAvroRegistryKafkaRecordDeserializationSchema.scala b/src/main/scala/io/epiphanous/flinkrunner/serde/ConfluentAvroRegistryKafkaRecordDeserializationSchema.scala
index 9fe9c95..57c0d5a 100644
--- a/src/main/scala/io/epiphanous/flinkrunner/serde/ConfluentAvroRegistryKafkaRecordDeserializationSchema.scala
+++ b/src/main/scala/io/epiphanous/flinkrunner/serde/ConfluentAvroRegistryKafkaRecordDeserializationSchema.scala
@@ -9,6 +9,7 @@ import io.epiphanous.flinkrunner.model.{
   FlinkEvent,
   KafkaSourceConfig
 }
+import org.apache.avro.specific.SpecificRecord
 import org.apache.flink.api.common.typeinfo.{TypeHint, TypeInformation}
 import org.apache.flink.connector.kafka.source.reader.deserializer.KafkaRecordDeserializationSchema
 import org.apache.flink.util.Collector
@@ -19,13 +20,19 @@ import java.util
 /**
  * A deserialization schema that uses the provided confluent avro schema
  * registry client and `fromKV` partial function to deserialize a kafka
- * key/value pair into an instance of a flink runner ADT.
+ * key/value pair into instances of a flink runner ADT.
  *
  * In order to decouple the shape of the flink runner ADT types from the
  * types that are serialized in kafka, a user of this class must provide a
  * `fromKV` partial function that maps from the specific key and value pair
- * (key is optional) deserialized from the kafka source into an instance of
- * the flink runner ADT.
+ * (key is optional) deserialized from the kafka source into a sequence of
+ * instances of the flink runner ADT. This means a single kafka record
+ * could generate multiple events.
+ *
+ * Note: that the number and size of the produced events should be
+ * relatively small. Depending on the source implementation records can be
+ * buffered in memory or collecting records might delay emitting the next
+ * checkpoint barrier.
  *
  * Usually, `fromKV` is as simple as providing a set of cases. Consider the
  * following example, where `A` and `B` are subclasses of the flink runner
@@ -34,11 +41,12 @@ import java.util
  * ignore the key and have defined our ADT types to be wrappers around the
  * deserialized records. However, you can use the deserialized key and
  * value in any way that makes sense for your application.
+ *
  * {{{
  *   {
  *     //    (key,value)     => ADT
- *     case (_, a:ASpecific) => A(a)
- *     case (_, b:BSpecific) => B(b)
+ *     case (_, a:ASpecific) => Seq(A(a))
+ *     case (_, b:BSpecific) => Seq(B(b))
  *   }
  * }}}
  * @param sourceName
@@ -48,18 +56,17 @@ import java.util
  * @param schemaRegistryClient
  *   the schema registry client
  * @param fromKV
- *   a partial function that should return a flink runner adt instance when
- *   passed a deserialized kafka key/value pair
+ *   a partial function that should return a sequence of zero or more flink
+ *   runner adt instances when passed a deserialized kafka key/value pair
  * @tparam ADT
  *   the flink runner ADT type
  */
 class ConfluentAvroRegistryKafkaRecordDeserializationSchema[
-    ADT <: FlinkEvent
-](
+    ADT <: FlinkEvent](
     sourceName: String,
     config: FlinkConfig[ADT],
     schemaRegistryClient: SchemaRegistryClient,
-    fromKV: PartialFunction[(Option[AnyRef], AnyRef), ADT]
+    fromKV: PartialFunction[(Option[AnyRef], AnyRef), Seq[ADT]]
 ) extends KafkaRecordDeserializationSchema[ADT]
     with LazyLogging {
 
@@ -95,7 +102,7 @@ class ConfluentAvroRegistryKafkaRecordDeserializationSchema[
     val key   =
       keyDeserializer.map(ds => ds.deserialize(topic, record.key()))
     val value = valueDeserializer.deserialize(topic, record.value())
-    if (Option(value).nonEmpty) out.collect(fromKV(key, value))
+    if (Option(value).nonEmpty) fromKV(key, value).map(out.collect)
   }
 
   override def getProducedType: TypeInformation[ADT] =