Internal duckdb#330: Quantile Performance Feedback

Restore the statistics API to use signed values.
Add coverage tests for quantile variants.

src/core_functions/aggregate/holistic/quantile.cpp

@@ -814,10 +814,10 @@ struct QuantileOperation {
 		const auto &stats = partition.stats;
 
 		//	If frames overlap significantly, then use local skip lists.
-		if (stats[0].end <= stats[1].start) {
+		if (stats[0].end <= stats[1].begin) {
 			//	Frames can overlap
-			const auto overlap = double(stats[1].start - stats[0].end);
-			const auto cover = double(stats[1].end - stats[0].start);
+			const auto overlap = double(stats[1].begin - stats[0].end);
+			const auto cover = double(stats[1].end - stats[0].begin);
 			const auto ratio = overlap / cover;
 			if (ratio > .75) {
 				return;

src/execution/window_executor.cpp

@@ -932,25 +932,25 @@ void WindowAggregateExecutor::Finalize() {
 	const int64_t count = aggregator->GetInputs().size();
 
 	//	First entry is the frame start
-	stats[0] = FrameBounds(-count, count);
+	stats[0] = FrameDelta(-count, count);
 	auto base = wexpr.expr_stats.empty() ? nullptr : wexpr.expr_stats[0].get();
 	switch (wexpr.start) {
 	case WindowBoundary::UNBOUNDED_PRECEDING:
 		stats[0].end = 0;
 		break;
 	case WindowBoundary::CURRENT_ROW_ROWS:
-		stats[0].start = stats[0].end = 0;
+		stats[0].begin = stats[0].end = 0;
 		break;
 	case WindowBoundary::EXPR_PRECEDING_ROWS:
 		if (base && base->GetStatsType() == StatisticsType::NUMERIC_STATS && NumericStats::HasMinMax(*base)) {
 			//	Preceding so negative offset from current row
-			stats[0].start = -NumericStats::GetMax<int64_t>(*base);
+			stats[0].begin = -NumericStats::GetMax<int64_t>(*base);
 			stats[0].end = -NumericStats::GetMin<int64_t>(*base) + 1;
 		}
 		break;
 	case WindowBoundary::EXPR_FOLLOWING_ROWS:
 		if (base && base->GetStatsType() == StatisticsType::NUMERIC_STATS && NumericStats::HasMinMax(*base)) {
-			stats[0].start = NumericStats::GetMin<int64_t>(*base);
+			stats[0].begin = NumericStats::GetMin<int64_t>(*base);
 			stats[0].end = NumericStats::GetMax<int64_t>(*base) + 1;
 		}
 		break;
@@ -964,25 +964,25 @@ void WindowAggregateExecutor::Finalize() {
 	}
 
 	//	Second entry is the frame end
-	stats[1] = FrameBounds(-count, count);
+	stats[1] = FrameDelta(-count, count);
 	base = wexpr.expr_stats.empty() ? nullptr : wexpr.expr_stats[1].get();
 	switch (wexpr.end) {
 	case WindowBoundary::UNBOUNDED_FOLLOWING:
-		stats[1].start = 0;
+		stats[1].begin = 0;
 		break;
 	case WindowBoundary::CURRENT_ROW_ROWS:
-		stats[1].start = stats[1].end = 0;
+		stats[1].begin = stats[1].end = 0;
 		break;
 	case WindowBoundary::EXPR_PRECEDING_ROWS:
 		if (base && base->GetStatsType() == StatisticsType::NUMERIC_STATS && NumericStats::HasMinMax(*base)) {
 			//	Preceding so negative offset from current row
-			stats[1].start = -NumericStats::GetMax<int64_t>(*base);
+			stats[1].begin = -NumericStats::GetMax<int64_t>(*base);
 			stats[1].end = -NumericStats::GetMin<int64_t>(*base) + 1;
 		}
 		break;
 	case WindowBoundary::EXPR_FOLLOWING_ROWS:
 		if (base && base->GetStatsType() == StatisticsType::NUMERIC_STATS && NumericStats::HasMinMax(*base)) {
-			stats[1].start = NumericStats::GetMin<int64_t>(*base);
+			stats[1].begin = NumericStats::GetMin<int64_t>(*base);
 			stats[1].end = NumericStats::GetMax<int64_t>(*base) + 1;
 		}
 		break;

src/include/duckdb/function/aggregate_function.hpp

@@ -16,8 +16,17 @@
 
 namespace duckdb {
 
+//! A half-open range of frame boundary values _relative to the current row_
+//! This is why they are signed values.
+struct FrameDelta {
+	FrameDelta() : begin(0), end(0) {};
+	FrameDelta(int64_t begin, int64_t end) : begin(begin), end(end) {};
+	int64_t begin = 0;
+	int64_t end = 0;
+};
+
 //! The half-open ranges of frame boundary values relative to the current row
-using FrameStats = array<FrameBounds, 2>;
+using FrameStats = array<FrameDelta, 2>;
 
 //! The partition data for custom window functions
 struct WindowPartitionInput {

test/sql/window/test_quantile_window.test_coverage

@@ -0,0 +1,80 @@
+# name: test/sql/window/test_quantile_window.test_coverage
+# description: Moving QUANTILE coverage, fixed or variable 100 element frame for MEDIAN, IQR, and MAD
+# group: [window]
+
+# Common table
+statement ok
+create table rank100 as
+    select b % 100 as a, b from range(10000000) tbl(b)
+
+# window_median_fixed_100
+query I
+select sum(m)
+from (
+    select median(a) over (
+        order by b asc
+        rows between 100 preceding and current row) as m
+    from rank100
+    ) q;
+----
+494997500
+
+# window_median_variable_100
+query I
+select sum(m)
+from (
+    select median(a) over (
+        order by b asc
+        rows between mod(b * 47, 521) preceding and 100 - mod(b * 47, 521) following) as m
+    from rank100
+    ) q;
+----
+494989867
+
+# window_iqr_fixed_100
+query II
+select min(iqr), max(iqr)
+from (
+    select quantile_cont(a, [0.25, 0.5, 0.75]) over (
+        order by b asc
+        rows between 100 preceding and current row) as iqr
+    from rank100
+    ) q;
+----
+[0.0, 0.0, 0.0]	[25.0, 50.0, 75.0]
+
+# window_iqr_variable_100
+query II
+select min(iqr), max(iqr)
+from (
+    select quantile_cont(a, [0.25, 0.5, 0.75]) over (
+        order by b asc
+        rows between mod(b * 47, 521) preceding and 100 - mod(b * 47, 521) following) as iqr
+    from rank100
+    ) q;
+----
+[0.0, 0.0, 0.0]	[76.5, 84.0, 91.5]
+
+# window_mad_fixed_100
+query I
+select sum(m)
+from (
+    select mad(a) over (
+        order by b asc
+        rows between 100 preceding and current row) as m
+    from rank100
+    ) q;
+----
+249998762.5
+
+# 
+query I
+select sum(m)
+from (
+    select mad(a) over (
+        order by b asc
+        rows between mod(b * 47, 521) preceding and 100 - mod(b * 47, 521) following) as m
+    from rank100
+    ) q;
+----
+249994596.000000