zsxwing commented on code in PR #40561:
URL: https://github.com/apache/spark/pull/40561#discussion_r1152772092
##########
sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/statefulOperators.scala:
##########
@@ -980,3 +1022,65 @@ object StreamingDeduplicateExec {
private val EMPTY_ROW =
UnsafeProjection.create(Array[DataType](NullType)).apply(InternalRow.apply(null))
}
+
+case class StreamingDeduplicateWithinWatermarkExec(
+ keyExpressions: Seq[Attribute],
+ child: SparkPlan,
+ stateInfo: Option[StatefulOperatorStateInfo] = None,
+ eventTimeWatermarkForLateEvents: Option[Long] = None,
+ eventTimeWatermarkForEviction: Option[Long] = None)
+ extends BaseStreamingDeduplicateExec {
+
+ protected val schemaForValueRow: StructType = StructType(
+ Array(StructField("expiresAt", LongType, nullable = false)))
+
+ protected val extraOptionOnStateStore: Map[String, String] = Map.empty
+
+ private val eventTimeCol: Attribute =
WatermarkSupport.findEventTimeColumn(child.output,
+ allowMultipleEventTimeColumns = false).get
+ private val delayThresholdMs =
eventTimeCol.metadata.getLong(EventTimeWatermark.delayKey)
+ private val eventTimeColOrdinal: Int = child.output.indexOf(eventTimeCol)
+
+ protected def initializeReusedDupInfoRow(): Option[UnsafeRow] = {
+ val timeoutToUnsafeRow = UnsafeProjection.create(schemaForValueRow)
+ val timeoutRow = timeoutToUnsafeRow(new
SpecificInternalRow(schemaForValueRow))
+ Some(timeoutRow)
+ }
+
+ protected def putDupInfoIntoState(
+ store: StateStore,
+ data: UnsafeRow,
+ key: UnsafeRow,
+ reusedDupInfoRow: Option[UnsafeRow]): Unit = {
+ assert(reusedDupInfoRow.isDefined, "This should have reused row.")
+ val timeoutRow = reusedDupInfoRow.get
+
+ val timestamp = data.getLong(eventTimeColOrdinal)
+ // The unit of timestamp in Spark is microseconds, convert the delay
threshold to micros.
+ val expiresAt = timestamp + delayThresholdMs * 1000
Review Comment:
nit: Should we use
`org.apache.spark.sql.catalyst.util.DateTimeUtils#millisToMicros` to handle
overflow (e.g., the user sets a very large `delayThresholdMs`)? We can fix this
later since you are just following the current `watermarkExpression` code.
##########
sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/statefulOperators.scala:
##########
@@ -980,3 +1022,65 @@ object StreamingDeduplicateExec {
private val EMPTY_ROW =
UnsafeProjection.create(Array[DataType](NullType)).apply(InternalRow.apply(null))
}
+
+case class StreamingDeduplicateWithinWatermarkExec(
+ keyExpressions: Seq[Attribute],
+ child: SparkPlan,
+ stateInfo: Option[StatefulOperatorStateInfo] = None,
+ eventTimeWatermarkForLateEvents: Option[Long] = None,
+ eventTimeWatermarkForEviction: Option[Long] = None)
+ extends BaseStreamingDeduplicateExec {
+
+ protected val schemaForValueRow: StructType = StructType(
+ Array(StructField("expiresAt", LongType, nullable = false)))
Review Comment:
QQ: why use LongType rather than `TimestampType`?
##########
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/Optimizer.scala:
##########
@@ -679,6 +679,8 @@ object RemoveNoopUnion extends Rule[LogicalPlan] {
d.withNewChildren(Seq(simplifyUnion(u)))
case d @ Deduplicate(_, u: Union) =>
d.withNewChildren(Seq(simplifyUnion(u)))
+ case d @ DeduplicateWithinWatermark(_, u: Union) =>
Review Comment:
> I don't think semantic deferences require disabling batch.
+1. I don't see why this should be different than the existing
`dropDuplicates`.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
