Examples with Operator org.apache.hadoop.hive.ql.exec.Operator used on opensource projects

Example 21 with Operator

use of org.apache.hadoop.hive.ql.exec.Operator in project hive by apache.

the class GenMRSkewJoinProcessor method processSkewJoin.

/**
 * Create tasks for processing skew joins. The idea is (HIVE-964) to use
 * separated jobs and map-joins to handle skew joins.
 * <p>
 * <ul>
 * <li>
 * Number of mr jobs to handle skew keys is the number of table minus 1 (we
 * can stream the last table, so big keys in the last table will not be a
 * problem).
 * <li>
 * At runtime in Join, we output big keys in one table into one corresponding
 * directories, and all same keys in other tables into different dirs(one for
 * each table). The directories will look like:
 * <ul>
 * <li>
 * dir-T1-bigkeys(containing big keys in T1), dir-T2-keys(containing keys
 * which is big in T1),dir-T3-keys(containing keys which is big in T1), ...
 * <li>
 * dir-T1-keys(containing keys which is big in T2), dir-T2-bigkeys(containing
 * big keys in T2),dir-T3-keys(containing keys which is big in T2), ...
 * <li>
 * dir-T1-keys(containing keys which is big in T3), dir-T2-keys(containing big
 * keys in T3),dir-T3-bigkeys(containing keys which is big in T3), ... .....
 * </ul>
 * </ul>
 * For each table, we launch one mapjoin job, taking the directory containing
 * big keys in this table and corresponding dirs in other tables as input.
 * (Actally one job for one row in the above.)
 *
 * <p>
 * For more discussions, please check
 * https://issues.apache.org/jira/browse/HIVE-964.
 */
@SuppressWarnings("unchecked")
public static void processSkewJoin(JoinOperator joinOp, Task<? extends Serializable> currTask, ParseContext parseCtx) throws SemanticException {
    // now does not work with outer joins
    if (!GenMRSkewJoinProcessor.skewJoinEnabled(parseCtx.getConf(), joinOp)) {
        return;
    }
    List<Task<? extends Serializable>> children = currTask.getChildTasks();
    Path baseTmpDir = parseCtx.getContext().getMRTmpPath();
    JoinDesc joinDescriptor = joinOp.getConf();
    Map<Byte, List<ExprNodeDesc>> joinValues = joinDescriptor.getExprs();
    int numAliases = joinValues.size();
    Map<Byte, Path> bigKeysDirMap = new HashMap<Byte, Path>();
    Map<Byte, Map<Byte, Path>> smallKeysDirMap = new HashMap<Byte, Map<Byte, Path>>();
    Map<Byte, Path> skewJoinJobResultsDir = new HashMap<Byte, Path>();
    Byte[] tags = joinDescriptor.getTagOrder();
    for (int i = 0; i < numAliases; i++) {
        Byte alias = tags[i];
        bigKeysDirMap.put(alias, getBigKeysDir(baseTmpDir, alias));
        Map<Byte, Path> smallKeysMap = new HashMap<Byte, Path>();
        smallKeysDirMap.put(alias, smallKeysMap);
        for (Byte src2 : tags) {
            if (!src2.equals(alias)) {
                smallKeysMap.put(src2, getSmallKeysDir(baseTmpDir, alias, src2));
            }
        }
        skewJoinJobResultsDir.put(alias, getBigKeysSkewJoinResultDir(baseTmpDir, alias));
    }
    joinDescriptor.setHandleSkewJoin(true);
    joinDescriptor.setBigKeysDirMap(bigKeysDirMap);
    joinDescriptor.setSmallKeysDirMap(smallKeysDirMap);
    joinDescriptor.setSkewKeyDefinition(HiveConf.getIntVar(parseCtx.getConf(), HiveConf.ConfVars.HIVESKEWJOINKEY));
    HashMap<Path, Task<? extends Serializable>> bigKeysDirToTaskMap = new HashMap<Path, Task<? extends Serializable>>();
    List<Serializable> listWorks = new ArrayList<Serializable>();
    List<Task<? extends Serializable>> listTasks = new ArrayList<Task<? extends Serializable>>();
    MapredWork currPlan = (MapredWork) currTask.getWork();
    TableDesc keyTblDesc = (TableDesc) currPlan.getReduceWork().getKeyDesc().clone();
    List<String> joinKeys = Utilities.getColumnNames(keyTblDesc.getProperties());
    List<String> joinKeyTypes = Utilities.getColumnTypes(keyTblDesc.getProperties());
    Map<Byte, TableDesc> tableDescList = new HashMap<Byte, TableDesc>();
    Map<Byte, RowSchema> rowSchemaList = new HashMap<Byte, RowSchema>();
    Map<Byte, List<ExprNodeDesc>> newJoinValues = new HashMap<Byte, List<ExprNodeDesc>>();
    Map<Byte, List<ExprNodeDesc>> newJoinKeys = new HashMap<Byte, List<ExprNodeDesc>>();
    // used for create mapJoinDesc, should be in order
    List<TableDesc> newJoinValueTblDesc = new ArrayList<TableDesc>();
    for (Byte tag : tags) {
        newJoinValueTblDesc.add(null);
    }
    for (int i = 0; i < numAliases; i++) {
        Byte alias = tags[i];
        List<ExprNodeDesc> valueCols = joinValues.get(alias);
        String colNames = "";
        String colTypes = "";
        int columnSize = valueCols.size();
        List<ExprNodeDesc> newValueExpr = new ArrayList<ExprNodeDesc>();
        List<ExprNodeDesc> newKeyExpr = new ArrayList<ExprNodeDesc>();
        ArrayList<ColumnInfo> columnInfos = new ArrayList<ColumnInfo>();
        boolean first = true;
        for (int k = 0; k < columnSize; k++) {
            TypeInfo type = valueCols.get(k).getTypeInfo();
            // any name, it does not matter.
            String newColName = i + "_VALUE_" + k;
            ColumnInfo columnInfo = new ColumnInfo(newColName, type, alias.toString(), false);
            columnInfos.add(columnInfo);
            newValueExpr.add(new ExprNodeColumnDesc(columnInfo));
            if (!first) {
                colNames = colNames + ",";
                colTypes = colTypes + ",";
            }
            first = false;
            colNames = colNames + newColName;
            colTypes = colTypes + valueCols.get(k).getTypeString();
        }
        // we are putting join keys at last part of the spilled table
        for (int k = 0; k < joinKeys.size(); k++) {
            if (!first) {
                colNames = colNames + ",";
                colTypes = colTypes + ",";
            }
            first = false;
            colNames = colNames + joinKeys.get(k);
            colTypes = colTypes + joinKeyTypes.get(k);
            ColumnInfo columnInfo = new ColumnInfo(joinKeys.get(k), TypeInfoFactory.getPrimitiveTypeInfo(joinKeyTypes.get(k)), alias.toString(), false);
            columnInfos.add(columnInfo);
            newKeyExpr.add(new ExprNodeColumnDesc(columnInfo));
        }
        newJoinValues.put(alias, newValueExpr);
        newJoinKeys.put(alias, newKeyExpr);
        tableDescList.put(alias, Utilities.getTableDesc(colNames, colTypes));
        rowSchemaList.put(alias, new RowSchema(columnInfos));
        // construct value table Desc
        String valueColNames = "";
        String valueColTypes = "";
        first = true;
        for (int k = 0; k < columnSize; k++) {
            // any name, it does not matter.
            String newColName = i + "_VALUE_" + k;
            if (!first) {
                valueColNames = valueColNames + ",";
                valueColTypes = valueColTypes + ",";
            }
            valueColNames = valueColNames + newColName;
            valueColTypes = valueColTypes + valueCols.get(k).getTypeString();
            first = false;
        }
        newJoinValueTblDesc.set(Byte.valueOf((byte) i), Utilities.getTableDesc(valueColNames, valueColTypes));
    }
    joinDescriptor.setSkewKeysValuesTables(tableDescList);
    joinDescriptor.setKeyTableDesc(keyTblDesc);
    for (int i = 0; i < numAliases - 1; i++) {
        Byte src = tags[i];
        MapWork newPlan = PlanUtils.getMapRedWork().getMapWork();
        // This code has been only added for testing
        boolean mapperCannotSpanPartns = parseCtx.getConf().getBoolVar(HiveConf.ConfVars.HIVE_MAPPER_CANNOT_SPAN_MULTIPLE_PARTITIONS);
        newPlan.setMapperCannotSpanPartns(mapperCannotSpanPartns);
        MapredWork clonePlan = SerializationUtilities.clonePlan(currPlan);
        Operator<? extends OperatorDesc>[] parentOps = new TableScanOperator[tags.length];
        for (int k = 0; k < tags.length; k++) {
            Operator<? extends OperatorDesc> ts = GenMapRedUtils.createTemporaryTableScanOperator(joinOp.getCompilationOpContext(), rowSchemaList.get((byte) k));
            ((TableScanOperator) ts).setTableDescSkewJoin(tableDescList.get((byte) k));
            parentOps[k] = ts;
        }
        Operator<? extends OperatorDesc> tblScan_op = parentOps[i];
        ArrayList<String> aliases = new ArrayList<String>();
        String alias = src.toString().intern();
        aliases.add(alias);
        Path bigKeyDirPath = bigKeysDirMap.get(src);
        newPlan.addPathToAlias(bigKeyDirPath, aliases);
        newPlan.getAliasToWork().put(alias, tblScan_op);
        PartitionDesc part = new PartitionDesc(tableDescList.get(src), null);
        newPlan.addPathToPartitionInfo(bigKeyDirPath, part);
        newPlan.getAliasToPartnInfo().put(alias, part);
        Operator<? extends OperatorDesc> reducer = clonePlan.getReduceWork().getReducer();
        assert reducer instanceof JoinOperator;
        JoinOperator cloneJoinOp = (JoinOperator) reducer;
        String dumpFilePrefix = "mapfile" + PlanUtils.getCountForMapJoinDumpFilePrefix();
        MapJoinDesc mapJoinDescriptor = new MapJoinDesc(newJoinKeys, keyTblDesc, newJoinValues, newJoinValueTblDesc, newJoinValueTblDesc, joinDescriptor.getOutputColumnNames(), i, joinDescriptor.getConds(), joinDescriptor.getFilters(), joinDescriptor.getNoOuterJoin(), dumpFilePrefix, joinDescriptor.getMemoryMonitorInfo(), joinDescriptor.getInMemoryDataSize());
        mapJoinDescriptor.setTagOrder(tags);
        mapJoinDescriptor.setHandleSkewJoin(false);
        mapJoinDescriptor.setNullSafes(joinDescriptor.getNullSafes());
        mapJoinDescriptor.setColumnExprMap(joinDescriptor.getColumnExprMap());
        MapredLocalWork localPlan = new MapredLocalWork(new LinkedHashMap<String, Operator<? extends OperatorDesc>>(), new LinkedHashMap<String, FetchWork>());
        Map<Byte, Path> smallTblDirs = smallKeysDirMap.get(src);
        for (int j = 0; j < numAliases; j++) {
            if (j == i) {
                continue;
            }
            Byte small_alias = tags[j];
            Operator<? extends OperatorDesc> tblScan_op2 = parentOps[j];
            localPlan.getAliasToWork().put(small_alias.toString(), tblScan_op2);
            Path tblDir = smallTblDirs.get(small_alias);
            localPlan.getAliasToFetchWork().put(small_alias.toString(), new FetchWork(tblDir, tableDescList.get(small_alias)));
        }
        newPlan.setMapRedLocalWork(localPlan);
        // construct a map join and set it as the child operator of tblScan_op
        MapJoinOperator mapJoinOp = (MapJoinOperator) OperatorFactory.getAndMakeChild(joinOp.getCompilationOpContext(), mapJoinDescriptor, (RowSchema) null, parentOps);
        // change the children of the original join operator to point to the map
        // join operator
        List<Operator<? extends OperatorDesc>> childOps = cloneJoinOp.getChildOperators();
        for (Operator<? extends OperatorDesc> childOp : childOps) {
            childOp.replaceParent(cloneJoinOp, mapJoinOp);
        }
        mapJoinOp.setChildOperators(childOps);
        HiveConf jc = new HiveConf(parseCtx.getConf(), GenMRSkewJoinProcessor.class);
        newPlan.setNumMapTasks(HiveConf.getIntVar(jc, HiveConf.ConfVars.HIVESKEWJOINMAPJOINNUMMAPTASK));
        newPlan.setMinSplitSize(HiveConf.getLongVar(jc, HiveConf.ConfVars.HIVESKEWJOINMAPJOINMINSPLIT));
        newPlan.setInputformat(HiveInputFormat.class.getName());
        MapredWork w = new MapredWork();
        w.setMapWork(newPlan);
        Task<? extends Serializable> skewJoinMapJoinTask = TaskFactory.get(w);
        skewJoinMapJoinTask.setFetchSource(currTask.isFetchSource());
        bigKeysDirToTaskMap.put(bigKeyDirPath, skewJoinMapJoinTask);
        listWorks.add(skewJoinMapJoinTask.getWork());
        listTasks.add(skewJoinMapJoinTask);
    }
    if (children != null) {
        for (Task<? extends Serializable> tsk : listTasks) {
            for (Task<? extends Serializable> oldChild : children) {
                tsk.addDependentTask(oldChild);
            }
        }
        currTask.setChildTasks(new ArrayList<Task<? extends Serializable>>());
        for (Task<? extends Serializable> oldChild : children) {
            oldChild.getParentTasks().remove(currTask);
        }
        listTasks.addAll(children);
    }
    ConditionalResolverSkewJoinCtx context = new ConditionalResolverSkewJoinCtx(bigKeysDirToTaskMap, children);
    ConditionalWork cndWork = new ConditionalWork(listWorks);
    ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork);
    cndTsk.setListTasks(listTasks);
    cndTsk.setResolver(new ConditionalResolverSkewJoin());
    cndTsk.setResolverCtx(context);
    currTask.setChildTasks(new ArrayList<Task<? extends Serializable>>());
    currTask.addDependentTask(cndTsk);
    return;
}

Example 22 with Operator

use of org.apache.hadoop.hive.ql.exec.Operator in project hive by apache.

the class GenMapRedUtils method splitTasks.

@SuppressWarnings("nls")
private static /**
 * Split two tasks by creating a temporary file between them.
 *
 * @param op reduce sink operator being processed
 * @param parentTask the parent task
 * @param childTask the child task
 * @param opProcCtx context
 */
void splitTasks(ReduceSinkOperator op, Task<? extends Serializable> parentTask, Task<? extends Serializable> childTask, GenMRProcContext opProcCtx) throws SemanticException {
    if (op.getNumParent() != 1) {
        throw new IllegalStateException("Expecting operator " + op + " to have one parent. " + "But found multiple parents : " + op.getParentOperators());
    }
    ParseContext parseCtx = opProcCtx.getParseCtx();
    parentTask.addDependentTask(childTask);
    // Root Task cannot depend on any other task, therefore childTask cannot be
    // a root Task
    List<Task<? extends Serializable>> rootTasks = opProcCtx.getRootTasks();
    if (rootTasks.contains(childTask)) {
        rootTasks.remove(childTask);
    }
    // Generate the temporary file name
    Context baseCtx = parseCtx.getContext();
    Path taskTmpDir = baseCtx.getMRTmpPath();
    Operator<? extends OperatorDesc> parent = op.getParentOperators().get(0);
    TableDesc tt_desc = PlanUtils.getIntermediateFileTableDesc(PlanUtils.getFieldSchemasFromRowSchema(parent.getSchema(), "temporarycol"));
    // Create the temporary file, its corresponding FileSinkOperaotr, and
    // its corresponding TableScanOperator.
    TableScanOperator tableScanOp = createTemporaryFile(parent, op, taskTmpDir, tt_desc, parseCtx);
    Map<Operator<? extends OperatorDesc>, GenMapRedCtx> mapCurrCtx = opProcCtx.getMapCurrCtx();
    mapCurrCtx.put(tableScanOp, new GenMapRedCtx(childTask, null));
    String streamDesc = taskTmpDir.toUri().toString();
    MapredWork cplan = (MapredWork) childTask.getWork();
    if (needsTagging(cplan.getReduceWork())) {
        Operator<? extends OperatorDesc> reducerOp = cplan.getReduceWork().getReducer();
        String id = null;
        if (reducerOp instanceof JoinOperator) {
            if (parseCtx.getJoinOps().contains(reducerOp)) {
                id = ((JoinOperator) reducerOp).getConf().getId();
            }
        } else if (reducerOp instanceof MapJoinOperator) {
            if (parseCtx.getMapJoinOps().contains(reducerOp)) {
                id = ((MapJoinOperator) reducerOp).getConf().getId();
            }
        } else if (reducerOp instanceof SMBMapJoinOperator) {
            if (parseCtx.getSmbMapJoinOps().contains(reducerOp)) {
                id = ((SMBMapJoinOperator) reducerOp).getConf().getId();
            }
        }
        if (id != null) {
            streamDesc = id + ":$INTNAME";
        } else {
            streamDesc = "$INTNAME";
        }
        String origStreamDesc = streamDesc;
        int pos = 0;
        while (cplan.getMapWork().getAliasToWork().get(streamDesc) != null) {
            streamDesc = origStreamDesc.concat(String.valueOf(++pos));
        }
        // TODO: Allocate work to remove the temporary files and make that
        // dependent on the redTask
        cplan.getReduceWork().setNeedsTagging(true);
    }
    // Add the path to alias mapping
    setTaskPlan(taskTmpDir, streamDesc, tableScanOp, cplan.getMapWork(), false, tt_desc);
    opProcCtx.setCurrTopOp(null);
    opProcCtx.setCurrAliasId(null);
    opProcCtx.setCurrTask(childTask);
    opProcCtx.addRootIfPossible(parentTask);
}

Example 23 with Operator

use of org.apache.hadoop.hive.ql.exec.Operator in project hive by apache.

the class SharedWorkOptimizer method areMergeable.

// FIXME: probably this should also be integrated with isSame() logics
private static boolean areMergeable(ParseContext pctx, SharedWorkOptimizerCache optimizerCache, TableScanOperator tsOp1, TableScanOperator tsOp2) throws SemanticException {
    // First we check if the two table scan operators can actually be merged
    // If schemas do not match, we currently do not merge
    List<String> prevTsOpNeededColumns = tsOp1.getNeededColumns();
    List<String> tsOpNeededColumns = tsOp2.getNeededColumns();
    if (prevTsOpNeededColumns.size() != tsOpNeededColumns.size()) {
        return false;
    }
    boolean notEqual = false;
    for (int i = 0; i < prevTsOpNeededColumns.size(); i++) {
        if (!prevTsOpNeededColumns.get(i).equals(tsOpNeededColumns.get(i))) {
            notEqual = true;
            break;
        }
    }
    if (notEqual) {
        return false;
    }
    // If row limit does not match, we currently do not merge
    if (tsOp1.getConf().getRowLimit() != tsOp2.getConf().getRowLimit()) {
        return false;
    }
    // If partitions do not match, we currently do not merge
    PrunedPartitionList prevTsOpPPList = pctx.getPrunedPartitions(tsOp1);
    PrunedPartitionList tsOpPPList = pctx.getPrunedPartitions(tsOp2);
    if (!prevTsOpPPList.getPartitions().equals(tsOpPPList.getPartitions())) {
        return false;
    }
    // If is a DPP, check if actually it refers to same target, column, etc.
    // Further, the DPP value needs to be generated from same subtree
    List<Operator<?>> dppsOp1 = new ArrayList<>(optimizerCache.tableScanToDPPSource.get(tsOp1));
    List<Operator<?>> dppsOp2 = new ArrayList<>(optimizerCache.tableScanToDPPSource.get(tsOp2));
    if (dppsOp1.isEmpty() && dppsOp2.isEmpty()) {
        return true;
    }
    for (int i = 0; i < dppsOp1.size(); i++) {
        Operator<?> op = dppsOp1.get(i);
        if (op instanceof ReduceSinkOperator) {
            Set<Operator<?>> ascendants = findAscendantWorkOperators(pctx, optimizerCache, op);
            if (ascendants.contains(tsOp2)) {
                // This should not happen, we cannot merge
                return false;
            }
        }
    }
    for (int i = 0; i < dppsOp2.size(); i++) {
        Operator<?> op = dppsOp2.get(i);
        if (op instanceof ReduceSinkOperator) {
            Set<Operator<?>> ascendants = findAscendantWorkOperators(pctx, optimizerCache, op);
            if (ascendants.contains(tsOp1)) {
                // This should not happen, we cannot merge
                return false;
            }
        }
    }
    if (dppsOp1.size() != dppsOp2.size()) {
        // Only first or second operator contains DPP pruning
        return false;
    }
    // Check if DPP branches are equal
    BitSet bs = new BitSet();
    for (int i = 0; i < dppsOp1.size(); i++) {
        Operator<?> dppOp1 = dppsOp1.get(i);
        for (int j = 0; j < dppsOp2.size(); j++) {
            if (!bs.get(j)) {
                // If not visited yet
                Operator<?> dppOp2 = dppsOp2.get(j);
                if (compareAndGatherOps(pctx, dppOp1, dppOp2) != null) {
                    // The DPP operator/branch are equal
                    bs.set(j);
                    break;
                }
            }
        }
        if (bs.cardinality() < i + 1) {
            return false;
        }
    }
    return true;
}

Example 24 with Operator

use of org.apache.hadoop.hive.ql.exec.Operator in project hive by apache.

the class SharedWorkOptimizer method pushFilterToTopOfTableScan.

private static void pushFilterToTopOfTableScan(SharedWorkOptimizerCache optimizerCache, TableScanOperator tsOp) throws UDFArgumentException {
    ExprNodeGenericFuncDesc tableScanExprNode = tsOp.getConf().getFilterExpr();
    List<Operator<? extends OperatorDesc>> allChildren = Lists.newArrayList(tsOp.getChildOperators());
    for (Operator<? extends OperatorDesc> op : allChildren) {
        if (op instanceof FilterOperator) {
            FilterOperator filterOp = (FilterOperator) op;
            ExprNodeDesc filterExprNode = filterOp.getConf().getPredicate();
            if (tableScanExprNode.isSame(filterExprNode)) {
                // We do not need to do anything
                return;
            }
            if (tableScanExprNode.getGenericUDF() instanceof GenericUDFOPOr) {
                for (ExprNodeDesc childExprNode : tableScanExprNode.getChildren()) {
                    if (childExprNode.isSame(filterExprNode)) {
                        // so probably we pushed previously
                        return;
                    }
                }
            }
            ExprNodeGenericFuncDesc newPred = ExprNodeGenericFuncDesc.newInstance(new GenericUDFOPAnd(), Arrays.<ExprNodeDesc>asList(tableScanExprNode.clone(), filterExprNode));
            filterOp.getConf().setPredicate(newPred);
        } else {
            Operator<FilterDesc> newOp = OperatorFactory.get(tsOp.getCompilationOpContext(), new FilterDesc(tableScanExprNode.clone(), false), new RowSchema(tsOp.getSchema().getSignature()));
            tsOp.replaceChild(op, newOp);
            newOp.getParentOperators().add(tsOp);
            op.replaceParent(tsOp, newOp);
            newOp.getChildOperators().add(op);
            // Add to cache (same group as tsOp)
            optimizerCache.putIfWorkExists(newOp, tsOp);
        }
    }
}

Example 25 with Operator

use of org.apache.hadoop.hive.ql.exec.Operator in project hive by apache.

the class SharedWorkOptimizer method extractSharedOptimizationInfoForRoot.

private static SharedResult extractSharedOptimizationInfoForRoot(ParseContext pctx, SharedWorkOptimizerCache optimizerCache, TableScanOperator retainableTsOp, TableScanOperator discardableTsOp) throws SemanticException {
    LinkedHashSet<Operator<?>> retainableOps = new LinkedHashSet<>();
    LinkedHashSet<Operator<?>> discardableOps = new LinkedHashSet<>();
    Set<Operator<?>> discardableInputOps = new HashSet<>();
    long dataSize = 0L;
    long maxDataSize = 0L;
    retainableOps.add(retainableTsOp);
    discardableOps.add(discardableTsOp);
    Operator<?> equalOp1 = retainableTsOp;
    Operator<?> equalOp2 = discardableTsOp;
    if (equalOp1.getNumChild() > 1 || equalOp2.getNumChild() > 1) {
        // TODO: Support checking multiple child operators to merge further.
        discardableInputOps.addAll(gatherDPPBranchOps(pctx, optimizerCache, discardableOps));
        return new SharedResult(retainableOps, discardableOps, discardableInputOps, dataSize, maxDataSize);
    }
    Operator<?> currentOp1 = retainableTsOp.getChildOperators().get(0);
    Operator<?> currentOp2 = discardableTsOp.getChildOperators().get(0);
    // Special treatment for Filter operator that ignores the DPP predicates
    if (currentOp1 instanceof FilterOperator && currentOp2 instanceof FilterOperator) {
        boolean equalFilters = false;
        FilterDesc op1Conf = ((FilterOperator) currentOp1).getConf();
        FilterDesc op2Conf = ((FilterOperator) currentOp2).getConf();
        if (op1Conf.getIsSamplingPred() == op2Conf.getIsSamplingPred() && StringUtils.equals(op1Conf.getSampleDescExpr(), op2Conf.getSampleDescExpr())) {
            Multiset<String> conjsOp1String = extractConjsIgnoringDPPPreds(op1Conf.getPredicate());
            Multiset<String> conjsOp2String = extractConjsIgnoringDPPPreds(op2Conf.getPredicate());
            if (conjsOp1String.equals(conjsOp2String)) {
                equalFilters = true;
            }
        }
        if (equalFilters) {
            equalOp1 = currentOp1;
            equalOp2 = currentOp2;
            retainableOps.add(equalOp1);
            discardableOps.add(equalOp2);
            if (currentOp1.getChildOperators().size() > 1 || currentOp2.getChildOperators().size() > 1) {
                // TODO: Support checking multiple child operators to merge further.
                discardableInputOps.addAll(gatherDPPBranchOps(pctx, optimizerCache, discardableOps));
                discardableInputOps.addAll(gatherDPPBranchOps(pctx, optimizerCache, retainableOps, discardableInputOps));
                return new SharedResult(retainableOps, discardableOps, discardableInputOps, dataSize, maxDataSize);
            }
            currentOp1 = currentOp1.getChildOperators().get(0);
            currentOp2 = currentOp2.getChildOperators().get(0);
        } else {
            // Bail out
            discardableInputOps.addAll(gatherDPPBranchOps(pctx, optimizerCache, discardableOps));
            discardableInputOps.addAll(gatherDPPBranchOps(pctx, optimizerCache, retainableOps, discardableInputOps));
            return new SharedResult(retainableOps, discardableOps, discardableInputOps, dataSize, maxDataSize);
        }
    }
    return extractSharedOptimizationInfo(pctx, optimizerCache, equalOp1, equalOp2, currentOp1, currentOp2, retainableOps, discardableOps, discardableInputOps, false);
}