[SPARK-38578][SQL] AdaptiveSparkPlanExec should ensure user-specified ordering

[ulysses-you]

What changes were proposed in this pull request?

• Ensure output ordering using requiredOrdering
• Override outputOrdering in AdaptiveSparkPlanExec

Why are the changes needed?

AdaptiveSparkPlanExec should ensure the output ordering is the requiredOrdering, so we leverage the EnsureRequirements to add sort if need.

FileFormatWriter will check and add an implicit sort for dynamic partition columns or bucket columns according to the input physical plan. The check became always failure since AQE AdaptiveSparkPlanExec has no outputOrdering.

That casues a redundant sort if user has specified a sort which satisfies the required ordering (dynamic partition and bucket columns).

Does this PR introduce any user-facing change?

no, improve performance

How was this patch tested?

add test

CREATETABLEt1 (c int) USING PARQUET PARTITIONED BY(p string);
CREATETABLEt2 USING PARQUET ASSELECT1 c, 'a' p;
INSERT INTO TABLE t1 PARTITION(p) select c, p from t2 order by p;

Before:

After:

[ulysses-you]

cc @maryannxue@cloud-fan

[cloud-fan]

do we need to override outputPartitioning as well?

[ulysses-you]

I think we can, but there is no requirements about outputPartitioning

[cloud-fan]

Seems we can remove this now?

[ulysses-you]

OptimizeSkewedJoin still use this. I considered unify them, but seems OptimizeSkewedJoin does not affect the required output ordering.

[cloud-fan]

We should refine the test name if we don't really test table insertion.

[ulysses-you]

yeah, refined it

[cloud-fan]

Suggested change

	// User-specified repartition is only effective when it's the root node, or under
	// User-specified sort is only effective when it's the root node, or under

[ulysses-you]

The failed test is irrelevant

ReportSinkMetricsSuite.test ReportSinkMetrics
org.scalatest.exceptions.TestFailedException: Expected null, but got {"metrics-1"="value-1", "metrics-2"="value-2"}

[cloud-fan]

Can we rerun the tests?

[ulysses-you]

rebased since that flaky test has been fixed

[cloud-fan]

After a second thought, what will happen if we just do a one-line fix override def outputOrdering: Seq[SortOrder] = requiredOrdering? AQE optimization may remove user-specified sort, but it doesn't matter as the file writing will add the necessary sort.

[ulysses-you]

file writing will add the necessary sort.

it does not. FileFormatWriter just checks the initialization SparkPlan if the ordering satisfies the required. If AQE changes the ordering at runtime, FileFormatWriter can not aware the change. So if we override the outputOrdering, we must make sure AQE won't change it.

[ulysses-you]

closed, in favor of #34568