[ 
https://issues.apache.org/jira/browse/DRILL-5796?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16542790#comment-16542790
 ] 

ASF GitHub Bot commented on DRILL-5796:
---------------------------------------

jbimbert commented on a change in pull request #1298: DRILL-5796: Filter 
pruning for multi rowgroup parquet file
URL: https://github.com/apache/drill/pull/1298#discussion_r202296895
 
 

 ##########
 File path: 
exec/java-exec/src/main/java/org/apache/drill/exec/expr/stat/ParquetIsPredicate.java
 ##########
 @@ -62,70 +62,73 @@ private ParquetIsPredicate(LogicalExpression expr, 
BiPredicate<Statistics<C>, Ra
     return visitor.visitUnknown(this, value);
   }
 
-  @Override
-  public boolean canDrop(RangeExprEvaluator<C> evaluator) {
+  /**
+   * Apply the filter condition against the meta of the rowgroup.
+   */
+  public RowsMatch matches(RangeExprEvaluator<C> evaluator) {
     Statistics<C> exprStat = expr.accept(evaluator, null);
-    if (isNullOrEmpty(exprStat)) {
-      return false;
-    }
+    return isNullOrEmpty(exprStat) ? RowsMatch.SOME : 
predicate.apply(exprStat, evaluator);
+  }
 
-    return predicate.test(exprStat, evaluator);
+  /**
+   * After the applying of the filter against the statistics of the rowgroup, 
if the result is RowsMatch.ALL,
+   * then we still must know if the rowgroup contains some null values, 
because it can change the filter result.
+   * If it contains some null values, then we change the RowsMatch.ALL into 
RowsMatch.SOME, which sya that maybe
+   * some values (the null ones) should be disgarded.
+   */
+  private static RowsMatch checkNull(Statistics exprStat) {
+    return hasNoNulls(exprStat) ? RowsMatch.ALL : RowsMatch.SOME;
   }
 
   /**
    * IS NULL predicate.
    */
   private static <C extends Comparable<C>> LogicalExpression 
createIsNullPredicate(LogicalExpression expr) {
     return new ParquetIsPredicate<C>(expr,
-        //if there are no nulls  -> canDrop
-        (exprStat, evaluator) -> hasNoNulls(exprStat)) {
-      private final boolean isArray = isArray(expr);
-
-      private boolean isArray(LogicalExpression expression) {
-        if (expression instanceof TypedFieldExpr) {
-          TypedFieldExpr typedFieldExpr = (TypedFieldExpr) expression;
-          SchemaPath schemaPath = typedFieldExpr.getPath();
-          return schemaPath.isArray();
-        }
-        return false;
-      }
-
-      @Override
-      public boolean canDrop(RangeExprEvaluator<C> evaluator) {
+      (exprStat, evaluator) -> {
         // for arrays we are not able to define exact number of nulls
         // [1,2,3] vs [1,2] -> in second case 3 is absent and thus it's null 
but statistics shows no nulls
-        return !isArray && super.canDrop(evaluator);
-      }
-    };
+        if (expr instanceof TypedFieldExpr) {
+          TypedFieldExpr typedFieldExpr = (TypedFieldExpr) expr;
+          if (typedFieldExpr.getPath().isArray()) {
+            return RowsMatch.SOME;
+          }
+        }
+        if (hasNoNulls(exprStat)) {
+          return RowsMatch.NONE;
+        }
+        return isAllNulls(exprStat, evaluator.getRowCount()) ? RowsMatch.ALL : 
RowsMatch.SOME;
+      });
   }
 
   /**
    * IS NOT NULL predicate.
    */
   private static <C extends Comparable<C>> LogicalExpression 
createIsNotNullPredicate(LogicalExpression expr) {
     return new ParquetIsPredicate<C>(expr,
-        //if there are all nulls  -> canDrop
-        (exprStat, evaluator) -> isAllNulls(exprStat, evaluator.getRowCount())
+      (exprStat, evaluator) -> isAllNulls(exprStat, evaluator.getRowCount()) ? 
RowsMatch.NONE : checkNull(exprStat)
     );
   }
 
   /**
    * IS TRUE predicate.
    */
   private static LogicalExpression createIsTruePredicate(LogicalExpression 
expr) {
-    return new ParquetIsPredicate<Boolean>(expr, (exprStat, evaluator) ->
-        //if max value is not true or if there are all nulls  -> canDrop
-        isAllNulls(exprStat, evaluator.getRowCount()) || 
exprStat.hasNonNullValue() && !((BooleanStatistics) exprStat).getMax()
-    );
+    return new ParquetIsPredicate<Boolean>(expr, (exprStat, evaluator) -> {
+      if (isAllNulls(exprStat, evaluator.getRowCount()) || 
exprStat.hasNonNullValue() && !((BooleanStatistics) exprStat).getMax()) {
+        return RowsMatch.NONE;
+      }
+      return ((BooleanStatistics) exprStat).getMin() ? checkNull(exprStat) : 
RowsMatch.SOME;
+    });
   }
 
   /**
    * IS FALSE predicate.
    */
   private static LogicalExpression createIsFalsePredicate(LogicalExpression 
expr) {
     return new ParquetIsPredicate<Boolean>(expr, (exprStat, evaluator) ->
-        //if min value is not false or if there are all nulls  -> canDrop
-        isAllNulls(exprStat, evaluator.getRowCount()) || 
exprStat.hasNonNullValue() && ((BooleanStatistics) exprStat).getMin()
+      exprStat.hasNonNullValue() && ((BooleanStatistics) exprStat).getMin() && 
((BooleanStatistics) exprStat).getMax() || isAllNulls(exprStat, 
evaluator.getRowCount()) ?
 
 Review comment:
   Added 4 new tests exec/java-exec/src/test/resources/parquetFilterPush/tfTbl/
   ff1 (false, false, 1 null)
   ft0 (false, true, 0 null)
   tt1 (true, true, 1 null)
   Other cases are already tested in 
exec/java-exec/src/test/resources/parquetFilterPush/blnTbl
   
   Updated "is true" and "is false" predicates.

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on 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


> Filter pruning for multi rowgroup parquet file
> ----------------------------------------------
>
>                 Key: DRILL-5796
>                 URL: https://issues.apache.org/jira/browse/DRILL-5796
>             Project: Apache Drill
>          Issue Type: Improvement
>          Components: Storage - Parquet
>            Reporter: Damien Profeta
>            Assignee: Jean-Blas IMBERT
>            Priority: Major
>             Fix For: 1.14.0
>
>
> Today, filter pruning use the file name as the partitioning key. This means 
> you can remove a partition only if the whole file is for the same partition. 
> With parquet, you can prune the filter if the rowgroup make a partition of 
> your dataset as the unit of work if the rowgroup not the file.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to