Examples with MapredLocalWork org.apache.hadoop.hive.ql.plan.MapredLocalWork used on opensource projects

Example 16 with MapredLocalWork

use of org.apache.hadoop.hive.ql.plan.MapredLocalWork in project hive by apache.

the class CommonJoinTaskDispatcher method mergeMapJoinTaskIntoItsChildMapRedTask.

/*
   * A task and its child task has been converted from join to mapjoin.
   * See if the two tasks can be merged.
   */
private void mergeMapJoinTaskIntoItsChildMapRedTask(MapRedTask mapJoinTask, Configuration conf) throws SemanticException {
    // If so, check if we can merge mapJoinTask into that child.
    if (mapJoinTask.getChildTasks() == null || mapJoinTask.getChildTasks().size() > 1) {
        // child-tasks in which case we don't want to do anything.
        return;
    }
    Task<?> childTask = mapJoinTask.getChildTasks().get(0);
    if (!(childTask instanceof MapRedTask)) {
        // Nothing to do if it is not a MapReduce task.
        return;
    }
    MapRedTask childMapRedTask = (MapRedTask) childTask;
    MapWork mapJoinMapWork = mapJoinTask.getWork().getMapWork();
    MapWork childMapWork = childMapRedTask.getWork().getMapWork();
    Map<String, Operator<? extends OperatorDesc>> mapJoinAliasToWork = mapJoinMapWork.getAliasToWork();
    if (mapJoinAliasToWork.size() > 1) {
        // Do not merge if the MapredWork of MapJoin has multiple input aliases.
        return;
    }
    Entry<String, Operator<? extends OperatorDesc>> mapJoinAliasToWorkEntry = mapJoinAliasToWork.entrySet().iterator().next();
    String mapJoinAlias = mapJoinAliasToWorkEntry.getKey();
    TableScanOperator mapJoinTaskTableScanOperator = OperatorUtils.findSingleOperator(mapJoinAliasToWorkEntry.getValue(), TableScanOperator.class);
    if (mapJoinTaskTableScanOperator == null) {
        throw new SemanticException("Expected a " + TableScanOperator.getOperatorName() + " operator as the work associated with alias " + mapJoinAlias + ". Found a " + mapJoinAliasToWork.get(mapJoinAlias).getName() + " operator.");
    }
    Set<FileSinkOperator> mapJoinTaskFileSinkOperators = OperatorUtils.findOperators(mapJoinTaskTableScanOperator, FileSinkOperator.class);
    if (mapJoinTaskFileSinkOperators.isEmpty()) {
        throw new SemanticException("Cannot find the " + FileSinkOperator.getOperatorName() + " operator at the last operator of the MapJoin Task.");
    }
    if (mapJoinTaskFileSinkOperators.size() > 1) {
        LOG.warn("Multiple " + FileSinkOperator.getOperatorName() + " operators found at the last operator of the MapJoin Task.");
        return;
    }
    // The mapJoinTaskFileSinkOperator writes to a different directory
    FileSinkOperator mapJoinTaskFileSinkOperator = mapJoinTaskFileSinkOperators.iterator().next();
    Path childMRPath = mapJoinTaskFileSinkOperator.getConf().getDirName();
    List<String> childMRAliases = childMapWork.getPathToAliases().get(childMRPath);
    if (childMRAliases == null || childMRAliases.size() != 1) {
        return;
    }
    String childMRAlias = childMRAliases.get(0);
    // Sanity check to make sure there is no alias conflict after merge.
    for (Entry<Path, List<String>> entry : childMapWork.getPathToAliases().entrySet()) {
        Path path = entry.getKey();
        List<String> aliases = entry.getValue();
        if (path.equals(childMRPath)) {
            continue;
        }
        if (aliases.contains(mapJoinAlias)) {
            // alias confict should not happen here.
            return;
        }
    }
    MapredLocalWork mapJoinLocalWork = mapJoinMapWork.getMapRedLocalWork();
    MapredLocalWork childLocalWork = childMapWork.getMapRedLocalWork();
    if ((mapJoinLocalWork != null && mapJoinLocalWork.getBucketMapjoinContext() != null) || (childLocalWork != null && childLocalWork.getBucketMapjoinContext() != null)) {
        // We should relax this constraint with a follow-up jira.
        return;
    }
    // is under the limit.
    if (!isLocalTableTotalSizeUnderLimitAfterMerge(conf, mapJoinLocalWork, childLocalWork)) {
        // Do not merge.
        return;
    }
    TableScanOperator childMRTaskTableScanOperator = OperatorUtils.findSingleOperator(childMapWork.getAliasToWork().get(childMRAlias.toString()), TableScanOperator.class);
    if (childMRTaskTableScanOperator == null) {
        throw new SemanticException("Expected a " + TableScanOperator.getOperatorName() + " operator as the work associated with alias " + childMRAlias + ". Found a " + childMapWork.getAliasToWork().get(childMRAlias).getName() + " operator.");
    }
    List<Operator<? extends OperatorDesc>> parentsInMapJoinTask = mapJoinTaskFileSinkOperator.getParentOperators();
    List<Operator<? extends OperatorDesc>> childrenInChildMRTask = childMRTaskTableScanOperator.getChildOperators();
    if (parentsInMapJoinTask.size() > 1 || childrenInChildMRTask.size() > 1) {
        // Do not merge if we do not know how to connect two operator trees.
        return;
    }
    // Step 2: Merge mapJoinTask into the Map-side of its child.
    // Step 2.1: Connect the operator trees of two MapRedTasks.
    Operator<? extends OperatorDesc> parentInMapJoinTask = parentsInMapJoinTask.get(0);
    Operator<? extends OperatorDesc> childInChildMRTask = childrenInChildMRTask.get(0);
    parentInMapJoinTask.replaceChild(mapJoinTaskFileSinkOperator, childInChildMRTask);
    childInChildMRTask.replaceParent(childMRTaskTableScanOperator, parentInMapJoinTask);
    // Step 2.2: Replace the corresponding part childMRWork's MapWork.
    GenMapRedUtils.replaceMapWork(mapJoinAlias, childMRAlias.toString(), mapJoinMapWork, childMapWork);
    // Step 2.3: Fill up stuff in local work
    if (mapJoinLocalWork != null) {
        if (childLocalWork == null) {
            childMapWork.setMapRedLocalWork(mapJoinLocalWork);
        } else {
            childLocalWork.getAliasToFetchWork().putAll(mapJoinLocalWork.getAliasToFetchWork());
            childLocalWork.getAliasToWork().putAll(mapJoinLocalWork.getAliasToWork());
        }
    }
    // Step 2.4: Remove this MapJoin task
    List<Task<?>> parentTasks = mapJoinTask.getParentTasks();
    mapJoinTask.setParentTasks(null);
    mapJoinTask.setChildTasks(null);
    childMapRedTask.getParentTasks().remove(mapJoinTask);
    if (parentTasks != null) {
        childMapRedTask.getParentTasks().addAll(parentTasks);
        for (Task<?> parentTask : parentTasks) {
            parentTask.getChildTasks().remove(mapJoinTask);
            if (!parentTask.getChildTasks().contains(childMapRedTask)) {
                parentTask.getChildTasks().add(childMapRedTask);
            }
        }
    } else {
        if (physicalContext.getRootTasks().contains(mapJoinTask)) {
            physicalContext.removeFromRootTask(mapJoinTask);
            if (childMapRedTask.getParentTasks() != null && childMapRedTask.getParentTasks().size() == 0 && !physicalContext.getRootTasks().contains(childMapRedTask)) {
                physicalContext.addToRootTask(childMapRedTask);
            }
        }
    }
    if (childMapRedTask.getParentTasks().size() == 0) {
        childMapRedTask.setParentTasks(null);
    }
}

Example 17 with MapredLocalWork

use of org.apache.hadoop.hive.ql.plan.MapredLocalWork 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.
 * (Actually 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<?> currTask, ParseContext parseCtx) throws SemanticException {
    // now does not work with outer joins
    if (!GenMRSkewJoinProcessor.skewJoinEnabled(parseCtx.getConf(), joinOp)) {
        return;
    }
    List<Task<?>> 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<?>> bigKeysDirToTaskMap = new HashMap<Path, Task<?>>();
    List<Serializable> listWorks = new ArrayList<Serializable>();
    List<Task<?>> listTasks = new ArrayList<Task<?>>();
    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<?> skewJoinMapJoinTask = TaskFactory.get(w);
        skewJoinMapJoinTask.setFetchSource(currTask.isFetchSource());
        bigKeysDirToTaskMap.put(bigKeyDirPath, skewJoinMapJoinTask);
        listWorks.add(skewJoinMapJoinTask.getWork());
        listTasks.add(skewJoinMapJoinTask);
    }
    if (children != null) {
        for (Task<?> tsk : listTasks) {
            for (Task<?> oldChild : children) {
                tsk.addDependentTask(oldChild);
            }
        }
        currTask.setChildTasks(new ArrayList<Task<?>>());
        for (Task<?> 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<?>>());
    currTask.addDependentTask(cndTsk);
    return;
}

Example 18 with MapredLocalWork

use of org.apache.hadoop.hive.ql.plan.MapredLocalWork in project hive by apache.

the class SparkSortMergeJoinFactory method setupBucketMapJoinInfo.

private static void setupBucketMapJoinInfo(MapWork plan, SMBMapJoinOperator currMapJoinOp) {
    if (currMapJoinOp != null) {
        Map<String, Map<String, List<String>>> aliasBucketFileNameMapping = currMapJoinOp.getConf().getAliasBucketFileNameMapping();
        if (aliasBucketFileNameMapping != null) {
            MapredLocalWork localPlan = plan.getMapRedLocalWork();
            if (localPlan == null) {
                localPlan = currMapJoinOp.getConf().getLocalWork();
            } else {
                // local plan is not null, we want to merge it into SMBMapJoinOperator's local work
                MapredLocalWork smbLocalWork = currMapJoinOp.getConf().getLocalWork();
                if (smbLocalWork != null) {
                    localPlan.getAliasToFetchWork().putAll(smbLocalWork.getAliasToFetchWork());
                    localPlan.getAliasToWork().putAll(smbLocalWork.getAliasToWork());
                }
            }
            if (localPlan == null) {
                return;
            }
            plan.setMapRedLocalWork(null);
            currMapJoinOp.getConf().setLocalWork(localPlan);
            BucketMapJoinContext bucketMJCxt = new BucketMapJoinContext();
            localPlan.setBucketMapjoinContext(bucketMJCxt);
            bucketMJCxt.setAliasBucketFileNameMapping(aliasBucketFileNameMapping);
            bucketMJCxt.setBucketFileNameMapping(currMapJoinOp.getConf().getBigTableBucketNumMapping());
            localPlan.setInputFileChangeSensitive(true);
            bucketMJCxt.setMapJoinBigTableAlias(currMapJoinOp.getConf().getBigTableAlias());
            bucketMJCxt.setBucketMatcherClass(org.apache.hadoop.hive.ql.exec.DefaultBucketMatcher.class);
            bucketMJCxt.setBigTablePartSpecToFileMapping(currMapJoinOp.getConf().getBigTablePartSpecToFileMapping());
            plan.setUseBucketizedHiveInputFormat(true);
        }
    }
}