HeartSaVioR commented on a change in pull request #27333:
URL: https://github.com/apache/spark/pull/27333#discussion_r511329896
##########
File path:
sql/core/src/test/scala/org/apache/spark/sql/streaming/FlatMapGroupsWithStateSuite.scala
##########
@@ -1019,6 +1019,56 @@ class FlatMapGroupsWithStateSuite extends
StateStoreMetricsTest {
spark.createDataset(Seq(("a", 2), ("b", 1))).toDF)
}
+ testWithAllStateVersions("SPARK-29438: ensure UNION doesn't lead
(flat)MapGroupsWithState" +
+ " to use shifted partition IDs") {
+ val stateFunc = (key: String, values: Iterator[String], state:
GroupState[RunningCount]) => {
+ val count = state.getOption.map(_.count).getOrElse(0L) + values.size
+ state.update(RunningCount(count))
+ (key, count.toString)
+ }
+
+ def constructUnionDf(desiredPartitionsForInput1: Int)
+ : (MemoryStream[String], MemoryStream[String], DataFrame) = {
+ val input1 = MemoryStream[String](desiredPartitionsForInput1)
+ val input2 = MemoryStream[String]
+ val df1 = input1.toDF()
+ .select($"value", $"value")
+ val df2 = input2.toDS()
+ .groupByKey(x => x)
+ .mapGroupsWithState(stateFunc) // Types = State: MyState, Out: (Str,
Str)
+ .toDF()
+
+ // Unioned DF would have columns as (String, String)
+ (input1, input2, df1.union(df2))
+ }
+
+ withTempDir { checkpointDir =>
+ val (input1, input2, unionDf) = constructUnionDf(2)
+ testStream(unionDf, Update)(
+ StartStream(checkpointLocation = checkpointDir.getAbsolutePath),
+ MultiAddData(input1, "input1-a")(input2, "input2-a"),
+ CheckNewAnswer(("input1-a", "input1-a"), ("input2-a", "1")),
+ StopStream
+ )
+
+ // We're restoring the query with different number of partitions in left
side of UNION,
+ // which may lead right side of union to have mismatched partition IDs
(e.g. if it relies on
+ // TaskContext.partitionId()). This test will verify
(flat)MapGroupsWithState doesn't have
+ // such issue.
+
+ val (newInput1, newInput2, newUnionDf) = constructUnionDf(3)
+
+ newInput1.addData("input1-a")
+ newInput2.addData("input2-a")
Review comment:
We're initializing the different memory streams here, which are not
aware of previous data and require us to fill the same data again on restore.
(Each addition to the memory stream represents a batch, hence you'll need to
add the same for the new memory stream.)
Probably there could be the another kind of small improvement for tests if
we have convenient way to push all the input rows we put before to the new
memory stream. That may be dealt with minor PR.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
