Mailing-List: contact dev-help@drill.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@drill.apache.org
From: jinfengni <git@git.apache.org>
To: dev@drill.apache.org
Reply-To: dev@drill.apache.org
References: <git-pr-882-drill@git.apache.org>
In-Reply-To: <git-pr-882-drill@git.apache.org>
Subject: [GitHub] drill pull request #882: DRILL-4735: ConvertCountToDirectScan rule enhanceme...
Content-Type: text/plain
Message-Id: <20170804180147.DC84FF334F@git1-us-west.apache.org>
Date: Fri,  4 Aug 2017 18:01:47 +0000 (UTC)
archived-at: Fri, 04 Aug 2017 18:01:50 -0000

Github user jinfengni commented on a diff in the pull request:

    https://github.com/apache/drill/pull/882#discussion_r131445381
  
    --- Diff: exec/java-exec/src/main/java/org/apache/drill/exec/planner/physical/ConvertCountToDirectScan.java ---
    @@ -85,109 +91,231 @@ protected ConvertCountToDirectScan(RelOptRuleOperand rule, String id) {
       @Override
       public void onMatch(RelOptRuleCall call) {
         final DrillAggregateRel agg = (DrillAggregateRel) call.rel(0);
    -    final DrillScanRel scan = (DrillScanRel) call.rel(call.rels.length -1);
    -    final DrillProjectRel proj = call.rels.length == 3 ? (DrillProjectRel) call.rel(1) : null;
    +    final DrillScanRel scan = (DrillScanRel) call.rel(call.rels.length - 1);
    +    final DrillProjectRel project = call.rels.length == 3 ? (DrillProjectRel) call.rel(1) : null;
     
         final GroupScan oldGrpScan = scan.getGroupScan();
         final PlannerSettings settings = PrelUtil.getPlannerSettings(call.getPlanner());
     
    -    // Only apply the rule when :
    +    // Only apply the rule when:
         //    1) scan knows the exact row count in getSize() call,
         //    2) No GroupBY key,
    -    //    3) only one agg function (Check if it's count(*) below).
    -    //    4) No distinct agg call.
    +    //    3) No distinct agg call.
         if (!(oldGrpScan.getScanStats(settings).getGroupScanProperty().hasExactRowCount()
             && agg.getGroupCount() == 0
    -        && agg.getAggCallList().size() == 1
             && !agg.containsDistinctCall())) {
           return;
         }
     
    -    AggregateCall aggCall = agg.getAggCallList().get(0);
    -
    -    if (aggCall.getAggregation().getName().equals("COUNT") ) {
    -
    -      long cnt = 0;
    -      //  count(*)  == >  empty arg  ==>  rowCount
    -      //  count(Not-null-input) ==> rowCount
    -      if (aggCall.getArgList().isEmpty() ||
    -          (aggCall.getArgList().size() == 1 &&
    -           ! agg.getInput().getRowType().getFieldList().get(aggCall.getArgList().get(0).intValue()).getType().isNullable())) {
    -        cnt = (long) oldGrpScan.getScanStats(settings).getRecordCount();
    -      } else if (aggCall.getArgList().size() == 1) {
    -      // count(columnName) ==> Agg ( Scan )) ==> columnValueCount
    -        int index = aggCall.getArgList().get(0);
    -
    -        if (proj != null) {
    -          // project in the middle of Agg and Scan : Only when input of AggCall is a RexInputRef in Project, we find the index of Scan's field.
    -          // For instance,
    -          // Agg - count($0)
    -          //  \
    -          //  Proj - Exp={$1}
    -          //    \
    -          //   Scan (col1, col2).
    -          // return count of "col2" in Scan's metadata, if found.
    -
    -          if (proj.getProjects().get(index) instanceof RexInputRef) {
    -            index = ((RexInputRef) proj.getProjects().get(index)).getIndex();
    -          } else {
    -            return;  // do not apply for all other cases.
    -          }
    -        }
    +    final CountsCollector countsCollector = new CountsCollector(settings);
    +    // if counts were not collected, rule won't be applied
    +    if (!countsCollector.collect(agg, scan, project)) {
    +      return;
    +    }
     
    -        String columnName = scan.getRowType().getFieldNames().get(index).toLowerCase();
    +    final RelDataType scanRowType = constructDataType(agg);
     
    -        cnt = oldGrpScan.getColumnValueCount(SchemaPath.getSimplePath(columnName));
    -        if (cnt == GroupScan.NO_COLUMN_STATS) {
    -          // if column stats are not available don't apply this rule
    -          return;
    -        }
    -      } else {
    -        return; // do nothing.
    -      }
    +    final DynamicPojoRecordReader<Long> reader = new DynamicPojoRecordReader<>(
    +        buildSchema(scanRowType.getFieldNames()),
    +        Collections.singletonList(countsCollector.getCounts()));
     
    -      RelDataType scanRowType = getCountDirectScanRowType(agg.getCluster().getTypeFactory());
    +    final ScanStats scanStats = new ScanStats(ScanStats.GroupScanProperty.EXACT_ROW_COUNT, 1, 1, scanRowType.getFieldCount());
    +    final GroupScan directScan = new MetadataDirectGroupScan(reader, oldGrpScan.getFiles(), scanStats);
     
    -      final ScanPrel newScan = ScanPrel.create(scan,
    -          scan.getTraitSet().plus(Prel.DRILL_PHYSICAL).plus(DrillDistributionTrait.SINGLETON), getCountDirectScan(cnt),
    -          scanRowType);
    +    final ScanPrel newScan = ScanPrel.create(scan,
    +        scan.getTraitSet().plus(Prel.DRILL_PHYSICAL).plus(DrillDistributionTrait.SINGLETON), directScan,
    +        scanRowType);
     
    -      List<RexNode> exprs = Lists.newArrayList();
    -      exprs.add(RexInputRef.of(0, scanRowType));
    +    final ProjectPrel newProject = new ProjectPrel(agg.getCluster(), agg.getTraitSet().plus(Prel.DRILL_PHYSICAL)
    +        .plus(DrillDistributionTrait.SINGLETON), newScan, prepareFieldExpressions(scanRowType), agg.getRowType());
     
    -      final ProjectPrel newProj = new ProjectPrel(agg.getCluster(), agg.getTraitSet().plus(Prel.DRILL_PHYSICAL)
    -          .plus(DrillDistributionTrait.SINGLETON), newScan, exprs, agg.getRowType());
    +    call.transformTo(newProject);
    +  }
     
    -      call.transformTo(newProj);
    +  /**
    +   * For each aggregate call creates field with "count$" prefix and bigint type.
    +   * Constructs record type for created fields.
    +   *
    +   * @param aggregateRel aggregate relation expression
    +   * @return record type
    +   */
    +  private RelDataType constructDataType(DrillAggregateRel aggregateRel) {
    +    List<RelDataTypeField> fields = new ArrayList<>();
    +    for (int i = 0; i < aggregateRel.getAggCallList().size(); i++) {
    +      RelDataTypeField field = new RelDataTypeFieldImpl("count$" + i, i, aggregateRel.getCluster().getTypeFactory().createSqlType(SqlTypeName.BIGINT));
    --- End diff --
    
    The fieldname in RelDataTypeField "count$0", "count$1" is not very informative. I think we could either get the field name from the aggregator operator, or use "count$colname" to indicate what column's count it represents. 


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastructure@apache.org or file a JIRA ticket
with INFRA.
---