squito commented on a change in pull request #28885:
URL: https://github.com/apache/spark/pull/28885#discussion_r460099447
##########
File path:
sql/core/src/test/scala/org/apache/spark/sql/execution/ExchangeSuite.scala
##########
@@ -156,4 +158,46 @@ class ExchangeSuite extends SparkPlanTest with
SharedSparkSession {
val projection2 = cached.select("_1", "_3").queryExecution.executedPlan
assert(!projection1.sameResult(projection2))
}
+
+ test("Exchange reuse across the whole plan") {
+ val df = sql(
+ """
+ |SELECT
+ | (SELECT max(a.key) FROM testData AS a JOIN testData AS b ON b.key =
a.key),
+ | a.key
+ |FROM testData AS a
+ |JOIN testData AS b ON b.key = a.key
+ """.stripMargin)
+
+ val plan = df.queryExecution.executedPlan
+
+ val exchangeIds = plan.collectWithSubqueries { case e: Exchange => e.id }
+ val reusedExchangeIds = plan.collectWithSubqueries {
+ case re: ReusedExchangeExec => re.child.id
+ }
+
+ assert(exchangeIds.size == 2, "Whole plan exchange reusing not working
correctly")
+ assert(reusedExchangeIds.size == 3, "Whole plan exchange reusing not
working correctly")
+ assert(reusedExchangeIds.forall(exchangeIds.contains(_)),
+ "ReusedExchangeExec should reuse an existing exchange")
+
+ val df2 = sql(
Review comment:
ah thanks, this helps a lot!
but, then isn't this a problem for *any* other rule which further modifies
the nodes, not just `ReuseSubquery`? It seems to solve this properly, you
actually want each `Exchange` to keep a reference to all the other exchanges
which will reuse it, which is preserved across the transformations. And then
some final pass which just resolves these links?
----------------------------------------------------------------
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]
