LuciferYang commented on code in PR #55422:
URL: https://github.com/apache/spark/pull/55422#discussion_r3122411364
##########
sql/core/src/main/scala/org/apache/spark/sql/execution/window/WindowEvaluatorFactoryBase.scala:
##########
@@ -295,4 +342,58 @@ trait WindowEvaluatorFactoryBase {
}
}
+ /**
+ * Segment-tree path eligibility. The tree relies on
+ * `DeclarativeAggregate.mergeExpressions`, which
[[AggregateWindowFunction]]s
+ * (NthValue, NTile, Rank, RowNumber, NullIndex) refuse via
+ * `mergeUnsupportedByWindowFunctionError`: they extend DeclarativeAggregate
+ * but are NOT merge-capable. Normal aggregate window expressions reach this
+ * code as the inner DeclarativeAggregate unwrapped from
+ * [[AggregateExpression]] (see `windowFrameExpressionFactoryPairs.collect`).
+ */
+ private def eligibleForSegTree(
+ functions: Array[Expression],
+ filters: Array[Option[Expression]],
+ frameType: FrameType): Boolean = {
+ // RANGE accepted only for single-column order specs. Multi-column RANGE
+ // with non-zero offset is already rejected by `createBoundOrdering`, so
+ // gating here on `orderSpec.size == 1` matches the Sliding-path invariant.
+ val frameTypeOk = frameType match {
+ case RowFrame => true
+ case RangeFrame => orderSpec.size == 1
+ }
+ SQLConf.get.windowSegmentTreeEnabled &&
+ frameTypeOk &&
+ filters.forall(_.isEmpty) &&
Review Comment:
`aggFilters` is built by pattern-matching `AggregateExpression.filter`. If
any analyzer rule rewrites `AGG(x) FILTER (WHERE p)` into `AGG(If(p, x, null))`
inside the aggregate function before Window, `ae.filter.isEmpty` returns true
and the filtered aggregate silently takes the segtree path. Should add a
defensive test (`AGG(x) FILTER (WHERE p) OVER (…)` with the conf enabled) that
asserts either (a) the fallback frame is used, or (b) results match the naive
oracle.
--
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]
