jcamachor commented on a change in pull request #567: HIVE-21382: Group by keys
reduction optimization - keys are not reduced in query23
URL: https://github.com/apache/hive/pull/567#discussion_r265249031
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/rules/HiveRelFieldTrimmer.java
##########
@@ -315,48 +322,128 @@ private boolean isRexLiteral(final RexNode rexNode) {
}
+ private static class TableRefFinder extends RexVisitorImpl<Void> {
+ private Set<RexTableInputRef> tableRefs = null;
+ TableRefFinder() {
+ super(true);
+ this.tableRefs = new HashSet<>();
+ }
+
+ public Set<RexTableInputRef> getTableRefs() {
+ return this.tableRefs;
+ }
+
+ @Override
+ public Void visitTableInputRef(RexTableInputRef ref) {
+ this.tableRefs.add(ref);
+ return null;
+ }
+ }
+
// Given a groupset this tries to find out if the cardinality of the
grouping columns could have changed
// because if not and it consist of keys (unique + not null OR pk), we can
safely remove rest of the columns
// if those are columns are not being used further up
private ImmutableBitSet generateGroupSetIfCardinalitySame(final Aggregate
aggregate,
final ImmutableBitSet
originalGroupSet, final ImmutableBitSet fieldsUsed) {
- Pair<RelOptTable, List<Integer>> tabToOrgCol =
HiveRelOptUtil.getColumnOriginSet(aggregate.getInput(),
-
originalGroupSet);
- if(tabToOrgCol == null) {
- return originalGroupSet;
- }
- RelOptHiveTable tbl = (RelOptHiveTable)tabToOrgCol.left;
- List<Integer> backtrackedGBList = tabToOrgCol.right;
- ImmutableBitSet backtrackedGBSet =
ImmutableBitSet.builder().addAll(backtrackedGBList).build();
- List<ImmutableBitSet> allKeys = tbl.getNonNullableKeys();
- ImmutableBitSet currentKey = null;
- for(ImmutableBitSet key:allKeys) {
- if(backtrackedGBSet.contains(key)) {
- // only if grouping sets consist of keys
- currentKey = key;
- break;
+ RexBuilder rexBuilder = aggregate.getCluster().getRexBuilder();
+ RelMetadataQuery mq = aggregate.getCluster().getMetadataQuery();
+
+ Iterator<Integer> iterator = originalGroupSet.iterator();
+ Map<Pair<RelOptTable, Integer>, Pair<List<Integer>, List<Integer>>>
mapGBKeysLineage= new HashMap<>();
+
+ Map<Pair<RelOptTable, Integer>, List<Integer>> candidateKeys = new
HashMap<>();
+
+ while(iterator.hasNext()) {
+ Integer key = iterator.next();
+ RexNode inputRef = rexBuilder.makeInputRef(aggregate.getInput(),
key.intValue());
+ Set<RexNode> exprLineage = mq.getExpressionLineage(aggregate, inputRef);
+ if(exprLineage != null && exprLineage.size() == 1){
+ RexNode expr = exprLineage.iterator().next();
+ if(expr instanceof RexTableInputRef) {
+ RexTableInputRef tblRef = (RexTableInputRef)expr;
+ Pair<RelOptTable, Integer> baseTable =
Pair.of(tblRef.getTableRef().getTable(),
tblRef.getTableRef().getEntityNumber());
+ if(mapGBKeysLineage.containsKey(baseTable)) {
+ List<Integer> baseCol = mapGBKeysLineage.get(baseTable).left;
+ baseCol.add(tblRef.getIndex());
+ List<Integer> gbKey = mapGBKeysLineage.get(baseTable).right;
+ gbKey.add(key);
+ } else {
+ List<Integer> baseCol = new ArrayList<>();
+ baseCol.add(tblRef.getIndex());
+ List<Integer> gbKey = new ArrayList<>();
+ gbKey.add(key);
+ mapGBKeysLineage.put(baseTable, Pair.of(baseCol, gbKey));
+ }
+ } else if(RexUtil.isDeterministic(expr)){
+ // even though we weren't able to backtrack this key it could still
be candidate for removal
+ // if rest of the columns contain pk/unique
+ TableRefFinder finder = new TableRefFinder();
Review comment:
We should use ```RexUtil.gatherTableReferences``` for same purpose so we do
not need to introduce ```TableRefFinder```.
----------------------------------------------------------------
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:
us...@infra.apache.org
With regards,
Apache Git Services
