vrozov 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_r202733948
########## File path: exec/java-exec/src/main/java/org/apache/drill/exec/expr/stat/ParquetIsPredicate.java ########## @@ -62,50 +62,51 @@ 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)) { Review comment: @arina-ielchiieva `num_nulls` must be -1 even for older files as well, otherwise, it is a bug in parquet library or drill that sets `num_nulls` incorrectly. If you have a concrete example where parquet library sets `num_nulls` to 0 where it should be -1, it is necessary to file JIRA for parquet. If drill sets `num_nulls` incorrectly while evaluating expressions or reading old parquet files, the bug needs to be fixed in drill. In the example you provided, should not statistics be empty (with `num_nulls` set to -1)? AFAIK, it is not valid to use `hasNonNullValue()` in place for nulls check as min/max may not be defined when all values are nulls or when a comparator is not defined for a data type. ---------------------------------------------------------------- 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 With regards, Apache Git Services