jcamachor commented on a change in pull request #1439:
URL: https://github.com/apache/hive/pull/1439#discussion_r479474959
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/rules/HiveAggregateJoinTransposeRule.java
##########
@@ -303,6 +305,90 @@ public void onMatch(RelOptRuleCall call) {
}
}
+ /**
+ * Determines weather the give grouping is unique.
+ *
+ * Consider a join which might produce non-unique rows; but later the
results are aggregated again.
+ * This method determines if there are sufficient columns in the grouping
which have been present previously as unique column(s).
+ */
+ private boolean isGroupingUnique(RelNode input, ImmutableBitSet groups) {
+ if (groups.isEmpty()) {
+ return false;
+ }
+ RelMetadataQuery mq = input.getCluster().getMetadataQuery();
+ Set<ImmutableBitSet> uKeys = mq.getUniqueKeys(input);
Review comment:
We could rely on `mq.areColumnsUnique`.
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/cost/HiveOnTezCostModel.java
##########
@@ -89,22 +89,23 @@ public RelOptCost getAggregateCost(HiveAggregate aggregate)
{
} else {
final RelMetadataQuery mq = aggregate.getCluster().getMetadataQuery();
// 1. Sum of input cardinalities
- final Double rCount = mq.getRowCount(aggregate.getInput());
- if (rCount == null) {
+ final Double inputRowCount = mq.getRowCount(aggregate.getInput());
+ final Double rowCount = mq.getRowCount(aggregate);
+ if (inputRowCount == null || rowCount == null) {
return null;
}
// 2. CPU cost = sorting cost
- final double cpuCost = algoUtils.computeSortCPUCost(rCount);
+ final double cpuCost = algoUtils.computeSortCPUCost(rowCount) +
inputRowCount * algoUtils.getCpuUnitCost();
Review comment:
Not sure about this change. If the algorithm is sort-based, you will
still sort the complete input, right?
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/rules/HiveAggregateJoinTransposeRule.java
##########
@@ -303,6 +305,90 @@ public void onMatch(RelOptRuleCall call) {
}
}
+ /**
+ * Determines weather the give grouping is unique.
+ *
+ * Consider a join which might produce non-unique rows; but later the
results are aggregated again.
+ * This method determines if there are sufficient columns in the grouping
which have been present previously as unique column(s).
+ */
+ private boolean isGroupingUnique(RelNode input, ImmutableBitSet groups) {
+ if (groups.isEmpty()) {
+ return false;
+ }
+ RelMetadataQuery mq = input.getCluster().getMetadataQuery();
+ Set<ImmutableBitSet> uKeys = mq.getUniqueKeys(input);
+ for (ImmutableBitSet u : uKeys) {
+ if (groups.contains(u)) {
+ return true;
+ }
+ }
+ if (input instanceof Join) {
+ Join join = (Join) input;
+ RexBuilder rexBuilder = input.getCluster().getRexBuilder();
+ SimpleConditionInfo cond = new SimpleConditionInfo(join.getCondition(),
rexBuilder);
+
+ if (cond.valid) {
+ ImmutableBitSet newGroup =
groups.intersect(ImmutableBitSet.fromBitSet(cond.fields));
+ RelNode l = join.getLeft();
+ RelNode r = join.getRight();
+
+ int joinFieldCount = join.getRowType().getFieldCount();
+ int lFieldCount = l.getRowType().getFieldCount();
+
+ ImmutableBitSet groupL = newGroup.get(0, lFieldCount);
+ ImmutableBitSet groupR = newGroup.get(lFieldCount,
joinFieldCount).shift(-lFieldCount);
+
+ if (isGroupingUnique(l, groupL)) {
Review comment:
This could call `mq.areColumnsUnique` instead of making the recursive
call.
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/rules/HiveAggregateJoinTransposeRule.java
##########
@@ -290,7 +291,8 @@ public void onMatch(RelOptRuleCall call) {
RelNode r = relBuilder.build();
RelOptCost afterCost = mq.getCumulativeCost(r);
RelOptCost beforeCost = mq.getCumulativeCost(aggregate);
- if (afterCost.isLt(beforeCost)) {
Review comment:
I think you suggested changing this... Maybe `isLe` if we do not
introduce an additional aggregate on top?
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/rules/HiveAggregateJoinTransposeRule.java
##########
@@ -290,7 +291,8 @@ public void onMatch(RelOptRuleCall call) {
RelNode r = relBuilder.build();
RelOptCost afterCost = mq.getCumulativeCost(r);
RelOptCost beforeCost = mq.getCumulativeCost(aggregate);
- if (afterCost.isLt(beforeCost)) {
+ boolean shouldForceTransform = isGroupingUnique(join,
aggregate.getGroupSet());
+ if (shouldForceTransform || afterCost.isLt(beforeCost)) {
Review comment:
Additionally, when we force triggering the transform, does it make sense
to verify that we are not creating an aggregate on top (i.e., we end up with
agg before join and after join?
That may narrow it down even further to a case when the pushdown should
always be beneficial.
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/rules/HiveAggregateJoinTransposeRule.java
##########
@@ -290,7 +291,8 @@ public void onMatch(RelOptRuleCall call) {
RelNode r = relBuilder.build();
RelOptCost afterCost = mq.getCumulativeCost(r);
RelOptCost beforeCost = mq.getCumulativeCost(aggregate);
- if (afterCost.isLt(beforeCost)) {
+ boolean shouldForceTransform = isGroupingUnique(join,
aggregate.getGroupSet());
Review comment:
The rule is enabled via config so we will only reach here if it is
enabled.
We should be able to force the transform even if cost-based variant is
disabled.
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/rules/HiveAggregateJoinTransposeRule.java
##########
@@ -303,6 +305,90 @@ public void onMatch(RelOptRuleCall call) {
}
}
+ /**
+ * Determines weather the give grouping is unique.
+ *
+ * Consider a join which might produce non-unique rows; but later the
results are aggregated again.
+ * This method determines if there are sufficient columns in the grouping
which have been present previously as unique column(s).
+ */
+ private boolean isGroupingUnique(RelNode input, ImmutableBitSet groups) {
+ if (groups.isEmpty()) {
+ return false;
+ }
+ RelMetadataQuery mq = input.getCluster().getMetadataQuery();
+ Set<ImmutableBitSet> uKeys = mq.getUniqueKeys(input);
+ for (ImmutableBitSet u : uKeys) {
+ if (groups.contains(u)) {
+ return true;
+ }
+ }
+ if (input instanceof Join) {
+ Join join = (Join) input;
+ RexBuilder rexBuilder = input.getCluster().getRexBuilder();
+ SimpleConditionInfo cond = new SimpleConditionInfo(join.getCondition(),
rexBuilder);
+
+ if (cond.valid) {
+ ImmutableBitSet newGroup =
groups.intersect(ImmutableBitSet.fromBitSet(cond.fields));
+ RelNode l = join.getLeft();
+ RelNode r = join.getRight();
+
+ int joinFieldCount = join.getRowType().getFieldCount();
+ int lFieldCount = l.getRowType().getFieldCount();
+
+ ImmutableBitSet groupL = newGroup.get(0, lFieldCount);
+ ImmutableBitSet groupR = newGroup.get(lFieldCount,
joinFieldCount).shift(-lFieldCount);
+
+ if (isGroupingUnique(l, groupL)) {
+ return true;
+ }
+ if (isGroupingUnique(r, groupR)) {
Review comment:
Same as above.
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/rules/HiveAggregateJoinTransposeRule.java
##########
@@ -303,6 +305,90 @@ public void onMatch(RelOptRuleCall call) {
}
}
+ /**
+ * Determines weather the give grouping is unique.
+ *
+ * Consider a join which might produce non-unique rows; but later the
results are aggregated again.
+ * This method determines if there are sufficient columns in the grouping
which have been present previously as unique column(s).
+ */
+ private boolean isGroupingUnique(RelNode input, ImmutableBitSet groups) {
+ if (groups.isEmpty()) {
+ return false;
+ }
+ RelMetadataQuery mq = input.getCluster().getMetadataQuery();
+ Set<ImmutableBitSet> uKeys = mq.getUniqueKeys(input);
+ for (ImmutableBitSet u : uKeys) {
+ if (groups.contains(u)) {
+ return true;
+ }
+ }
+ if (input instanceof Join) {
+ Join join = (Join) input;
+ RexBuilder rexBuilder = input.getCluster().getRexBuilder();
+ SimpleConditionInfo cond = new SimpleConditionInfo(join.getCondition(),
rexBuilder);
+
+ if (cond.valid) {
+ ImmutableBitSet newGroup =
groups.intersect(ImmutableBitSet.fromBitSet(cond.fields));
+ RelNode l = join.getLeft();
+ RelNode r = join.getRight();
+
+ int joinFieldCount = join.getRowType().getFieldCount();
+ int lFieldCount = l.getRowType().getFieldCount();
+
+ ImmutableBitSet groupL = newGroup.get(0, lFieldCount);
+ ImmutableBitSet groupR = newGroup.get(lFieldCount,
joinFieldCount).shift(-lFieldCount);
+
+ if (isGroupingUnique(l, groupL)) {
+ return true;
+ }
+ if (isGroupingUnique(r, groupR)) {
+ return true;
+ }
+ }
+ }
+ return false;
+ }
+
+ static class SimpleConditionInfo {
Review comment:
Can you use `JoinInfo.of` for this? It seems it does something very
similar.
----------------------------------------------------------------
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]
