hsyuan commented on a change in pull request #2006:
URL: https://github.com/apache/calcite/pull/2006#discussion_r437062358



##########
File path: 
core/src/main/java/org/apache/calcite/adapter/enumerable/EnumerableMergeJoin.java
##########
@@ -194,6 +250,81 @@ public static boolean isMergeJoinSupported(JoinRelType 
joinType) {
     return DeriveMode.BOTH;
   }
 
+  private List<Integer> immutableIntListToList(ImmutableIntList intList, int 
offset) {

Review comment:
       Can you rebase on master? I just updated this file.
   Use `ImmutableIntList.incr` instead.

##########
File path: 
core/src/main/java/org/apache/calcite/adapter/enumerable/EnumerableMergeJoin.java
##########
@@ -194,6 +250,81 @@ public static boolean isMergeJoinSupported(JoinRelType 
joinType) {
     return DeriveMode.BOTH;
   }
 
+  private List<Integer> immutableIntListToList(ImmutableIntList intList, int 
offset) {
+    ArrayList<Integer> arrayList = new ArrayList<>(intList.size());
+    for (int i : intList) {
+      arrayList.add(i + offset);
+    }
+    return arrayList;
+  }
+
+  private boolean isSubset(List<Integer> a, List<Integer> b) {
+    if (a.size() > b.size()) {
+      return false;
+    }
+    Set<Integer> set = new HashSet<>(b);
+    for (int i = 0; i < a.size(); i++) {
+      if (!set.contains(a.get(i))) {
+        return false;
+      }
+    }
+    return true;
+  }
+
+  private boolean isPrefixOrderingNotRequired(List<Integer> a, List<Integer> 
b) {

Review comment:
       See `RelCollations.containsOrderless`, you can make the private one 
public is necessary.

##########
File path: 
core/src/main/java/org/apache/calcite/adapter/enumerable/EnumerableMergeJoin.java
##########
@@ -101,28 +103,55 @@ public static boolean isMergeJoinSupported(JoinRelType 
joinType) {
     // Required collation keys can be subset or superset of merge join keys.
     RelCollation collation = required.getTrait(RelCollationTraitDef.INSTANCE);
     List<Integer> reqKeys = RelCollations.ordinals(collation);
-    ImmutableBitSet reqKeySet = ImmutableBitSet.of(reqKeys);
+    // need to copy reqKeys because it is not sortable.
+    reqKeys = new ArrayList<>(reqKeys);
+    List<Integer> leftKeys = immutableIntListToList(joinInfo.leftKeys, 0);
+    List<Integer> rightKeys = immutableIntListToList(joinInfo.rightKeys,
+        left.getRowType().getFieldCount());
+    List<Integer> rightKeysNotShifted = 
immutableIntListToList(joinInfo.rightKeys, 0);
 
+    ImmutableBitSet reqKeySet = ImmutableBitSet.of(reqKeys);
     ImmutableBitSet leftKeySet = ImmutableBitSet.of(joinInfo.leftKeys);
     ImmutableBitSet rightKeySet = ImmutableBitSet.of(joinInfo.rightKeys)
         .shift(left.getRowType().getFieldCount());
 
     Map<Integer, Integer> keyMap = new HashMap<>();
-    final int keyCount = leftKeySet.cardinality();
+    final int keyCount = leftKeys.size();

Review comment:
       nice catch. in case the keys are `1,1,2,2`. IMHO, we shouldn't see 
duplicate join keys in physical merge/hash join operators. They should be 
optimized away, because that means the predicate is not pushed down at all.
   like foo.a = bar.b and foo.a=bar.c. The predicate bar.b=bar.c should be 
pushed down for table bar.

##########
File path: 
core/src/main/java/org/apache/calcite/adapter/enumerable/EnumerableMergeJoin.java
##########
@@ -101,28 +103,55 @@ public static boolean isMergeJoinSupported(JoinRelType 
joinType) {
     // Required collation keys can be subset or superset of merge join keys.
     RelCollation collation = required.getTrait(RelCollationTraitDef.INSTANCE);
     List<Integer> reqKeys = RelCollations.ordinals(collation);
-    ImmutableBitSet reqKeySet = ImmutableBitSet.of(reqKeys);
+    // need to copy reqKeys because it is not sortable.
+    reqKeys = new ArrayList<>(reqKeys);

Review comment:
       why do you want to sort it?

##########
File path: 
core/src/main/java/org/apache/calcite/adapter/enumerable/EnumerableMergeJoin.java
##########
@@ -101,28 +103,55 @@ public static boolean isMergeJoinSupported(JoinRelType 
joinType) {
     // Required collation keys can be subset or superset of merge join keys.
     RelCollation collation = required.getTrait(RelCollationTraitDef.INSTANCE);
     List<Integer> reqKeys = RelCollations.ordinals(collation);
-    ImmutableBitSet reqKeySet = ImmutableBitSet.of(reqKeys);
+    // need to copy reqKeys because it is not sortable.
+    reqKeys = new ArrayList<>(reqKeys);
+    List<Integer> leftKeys = immutableIntListToList(joinInfo.leftKeys, 0);
+    List<Integer> rightKeys = immutableIntListToList(joinInfo.rightKeys,
+        left.getRowType().getFieldCount());
+    List<Integer> rightKeysNotShifted = 
immutableIntListToList(joinInfo.rightKeys, 0);
 
+    ImmutableBitSet reqKeySet = ImmutableBitSet.of(reqKeys);
     ImmutableBitSet leftKeySet = ImmutableBitSet.of(joinInfo.leftKeys);
     ImmutableBitSet rightKeySet = ImmutableBitSet.of(joinInfo.rightKeys)
         .shift(left.getRowType().getFieldCount());
 
     Map<Integer, Integer> keyMap = new HashMap<>();
-    final int keyCount = leftKeySet.cardinality();
+    final int keyCount = leftKeys.size();
     for (int i = 0; i < keyCount; i++) {
       keyMap.put(joinInfo.leftKeys.get(i), joinInfo.rightKeys.get(i));
     }
     Mappings.TargetMapping mapping = Mappings.target(keyMap,
         left.getRowType().getFieldCount(),
         right.getRowType().getFieldCount());
 
-    // Only consider exact key match for now
+
     if (reqKeySet.equals(leftKeySet)) {
+      // if sort keys equal to left join keys, we can pass through all 
collations directly.
       RelCollation rightCollation = RexUtil.apply(mapping, collation);
       return Pair.of(
           required, ImmutableList.of(required,
           required.replace(rightCollation)));
+    } else if (isSubset(reqKeys, leftKeys)) {
+      // if sort keys are subset of left join keys, we can extend collations 
to make sure all join
+      // keys are sorted.
+      collation = extendCollation(collation, leftKeys);
+      RelCollation rightCollation = RexUtil.apply(mapping, collation);
+      return Pair.of(
+          required, ImmutableList.of(required.replace(collation),
+              required.replace(rightCollation)));
+    } else if (isPrefixOrderingNotRequired(leftKeys, reqKeys)

Review comment:
       You can use RelCollations.containsOrderless




----------------------------------------------------------------
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]


Reply via email to