Search in sources :

Example 1 with ConditionalResolverCommonJoin

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

the class SortMergeJoinTaskDispatcher method processCurrentTask.

@Override
public Task<?> processCurrentTask(MapRedTask currTask, ConditionalTask conditionalTask, Context context) throws SemanticException {
    // whether it contains a sort merge join operator
    MapredWork currWork = currTask.getWork();
    SMBMapJoinOperator originalSMBJoinOp = getSMBMapJoinOp(currWork);
    if (!isEligibleForOptimization(originalSMBJoinOp)) {
        return null;
    }
    currTask.setTaskTag(Task.CONVERTED_SORTMERGEJOIN);
    // Convert the work containing to sort-merge join into a work, as if it had a regular join.
    // Note that the operator tree is not changed - is still contains the SMB join, but the
    // plan is changed (aliasToWork etc.) to contain all the paths as if it was a regular join.
    // This is used to convert the plan to a map-join, and then the original SMB join plan is used
    // as a backup task.
    MapredWork currJoinWork = convertSMBWorkToJoinWork(currWork, originalSMBJoinOp);
    SMBMapJoinOperator newSMBJoinOp = getSMBMapJoinOp(currJoinWork);
    currWork.getMapWork().setLeftInputJoin(originalSMBJoinOp.getConf().isLeftInputJoin());
    currWork.getMapWork().setBaseSrc(originalSMBJoinOp.getConf().getBaseSrc());
    currWork.getMapWork().setMapAliases(originalSMBJoinOp.getConf().getMapAliases());
    currJoinWork.getMapWork().setLeftInputJoin(originalSMBJoinOp.getConf().isLeftInputJoin());
    currJoinWork.getMapWork().setBaseSrc(originalSMBJoinOp.getConf().getBaseSrc());
    currJoinWork.getMapWork().setMapAliases(originalSMBJoinOp.getConf().getMapAliases());
    // create conditional work list and task list
    List<Serializable> listWorks = new ArrayList<Serializable>();
    List<Task<?>> listTasks = new ArrayList<Task<?>>();
    // create task to aliases mapping and alias to input file mapping for resolver
    // Must be deterministic order map for consistent q-test output across Java versions
    HashMap<Task<?>, Set<String>> taskToAliases = new LinkedHashMap<Task<?>, Set<String>>();
    // Note that pathToAlias will behave as if the original plan was a join plan
    Map<Path, List<String>> pathToAliases = currJoinWork.getMapWork().getPathToAliases();
    // generate a map join task for the big table
    SMBJoinDesc originalSMBJoinDesc = originalSMBJoinOp.getConf();
    Byte[] order = originalSMBJoinDesc.getTagOrder();
    int numAliases = order.length;
    Set<Integer> bigTableCandidates = MapJoinProcessor.getBigTableCandidates(originalSMBJoinDesc.getConds());
    HashMap<String, Long> aliasToSize = new HashMap<String, Long>();
    Configuration conf = context.getConf();
    try {
        long aliasTotalKnownInputSize = getTotalKnownInputSize(context, currJoinWork.getMapWork(), pathToAliases, aliasToSize);
        long ThresholdOfSmallTblSizeSum = HiveConf.getLongVar(conf, HiveConf.ConfVars.HIVESMALLTABLESFILESIZE);
        for (int bigTablePosition = 0; bigTablePosition < numAliases; bigTablePosition++) {
            // this table cannot be big table
            if (!bigTableCandidates.contains(bigTablePosition)) {
                continue;
            }
            // create map join task for the given big table position
            MapRedTask newTask = convertSMBTaskToMapJoinTask(currJoinWork, bigTablePosition, newSMBJoinOp);
            MapWork mapWork = newTask.getWork().getMapWork();
            Operator<?> parentOp = originalSMBJoinOp.getParentOperators().get(bigTablePosition);
            Set<String> aliases = GenMapRedUtils.findAliases(mapWork, parentOp);
            long aliasKnownSize = Utilities.sumOf(aliasToSize, aliases);
            if (aliasKnownSize > 0) {
                long smallTblTotalKnownSize = aliasTotalKnownInputSize - aliasKnownSize;
                if (smallTblTotalKnownSize > ThresholdOfSmallTblSizeSum) {
                    // this table is not good to be a big table.
                    continue;
                }
            }
            // add into conditional task
            listWorks.add(newTask.getWork());
            listTasks.add(newTask);
            newTask.setTaskTag(Task.CONVERTED_MAPJOIN);
            newTask.setFetchSource(currTask.isFetchSource());
            // set up backup task
            newTask.setBackupTask(currTask);
            newTask.setBackupChildrenTasks(currTask.getChildTasks());
            // put the mapping task to aliases
            taskToAliases.put(newTask, aliases);
        }
    } catch (Exception e) {
        throw new SemanticException("Generate Map Join Task Error", e);
    }
    // insert current common join task to conditional task
    listWorks.add(currTask.getWork());
    listTasks.add(currTask);
    // clear JoinTree and OP Parse Context
    currWork.getMapWork().setLeftInputJoin(false);
    currWork.getMapWork().setBaseSrc(null);
    currWork.getMapWork().setMapAliases(null);
    // create conditional task and insert conditional task into task tree
    ConditionalWork cndWork = new ConditionalWork(listWorks);
    ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork);
    cndTsk.setListTasks(listTasks);
    // set resolver and resolver context
    cndTsk.setResolver(new ConditionalResolverCommonJoin());
    ConditionalResolverCommonJoinCtx resolverCtx = new ConditionalResolverCommonJoinCtx();
    resolverCtx.setPathToAliases(pathToAliases);
    resolverCtx.setAliasToKnownSize(aliasToSize);
    resolverCtx.setTaskToAliases(taskToAliases);
    resolverCtx.setCommonJoinTask(currTask);
    resolverCtx.setLocalTmpDir(context.getLocalScratchDir(false));
    resolverCtx.setHdfsTmpDir(context.getMRScratchDir());
    cndTsk.setResolverCtx(resolverCtx);
    // replace the current task with the new generated conditional task
    replaceTaskWithConditionalTask(currTask, cndTsk);
    return cndTsk;
}
Also used : Serializable(java.io.Serializable) ConditionalTask(org.apache.hadoop.hive.ql.exec.ConditionalTask) Task(org.apache.hadoop.hive.ql.exec.Task) MapRedTask(org.apache.hadoop.hive.ql.exec.mr.MapRedTask) Set(java.util.Set) SMBJoinDesc(org.apache.hadoop.hive.ql.plan.SMBJoinDesc) Configuration(org.apache.hadoop.conf.Configuration) HashMap(java.util.HashMap) LinkedHashMap(java.util.LinkedHashMap) ArrayList(java.util.ArrayList) ConditionalWork(org.apache.hadoop.hive.ql.plan.ConditionalWork) SMBMapJoinOperator(org.apache.hadoop.hive.ql.exec.SMBMapJoinOperator) LinkedHashMap(java.util.LinkedHashMap) MapRedTask(org.apache.hadoop.hive.ql.exec.mr.MapRedTask) ConditionalResolverCommonJoinCtx(org.apache.hadoop.hive.ql.plan.ConditionalResolverCommonJoin.ConditionalResolverCommonJoinCtx) MapredWork(org.apache.hadoop.hive.ql.plan.MapredWork) ConditionalTask(org.apache.hadoop.hive.ql.exec.ConditionalTask) ArrayList(java.util.ArrayList) List(java.util.List) ConditionalResolverCommonJoin(org.apache.hadoop.hive.ql.plan.ConditionalResolverCommonJoin) SemanticException(org.apache.hadoop.hive.ql.parse.SemanticException) Path(org.apache.hadoop.fs.Path) SemanticException(org.apache.hadoop.hive.ql.parse.SemanticException) MapWork(org.apache.hadoop.hive.ql.plan.MapWork)

Example 2 with ConditionalResolverCommonJoin

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

the class CommonJoinTaskDispatcher method processCurrentTask.

@Override
public Task<?> processCurrentTask(MapRedTask currTask, ConditionalTask conditionalTask, Context context) throws SemanticException {
    // whether it contains common join op; if contains, return this common join op
    JoinOperator joinOp = getJoinOp(currTask);
    if (joinOp == null || joinOp.getConf().isFixedAsSorted()) {
        return null;
    }
    currTask.setTaskTag(Task.COMMON_JOIN);
    MapWork currWork = currTask.getWork().getMapWork();
    // create conditional work list and task list
    List<Serializable> listWorks = new ArrayList<Serializable>();
    List<Task<?>> listTasks = new ArrayList<Task<?>>();
    // create task to aliases mapping and alias to input file mapping for resolver
    // Must be deterministic order map for consistent q-test output across Java versions
    HashMap<Task<?>, Set<String>> taskToAliases = new LinkedHashMap<Task<?>, Set<String>>();
    Map<Path, List<String>> pathToAliases = currWork.getPathToAliases();
    Map<String, Operator<? extends OperatorDesc>> aliasToWork = currWork.getAliasToWork();
    // start to generate multiple map join tasks
    JoinDesc joinDesc = joinOp.getConf();
    if (aliasToSize == null) {
        aliasToSize = new HashMap<String, Long>();
    }
    try {
        long aliasTotalKnownInputSize = getTotalKnownInputSize(context, currWork, pathToAliases, aliasToSize);
        Set<Integer> bigTableCandidates = MapJoinProcessor.getBigTableCandidates(joinDesc.getConds());
        // no table could be the big table; there is no need to convert
        if (bigTableCandidates.isEmpty()) {
            return null;
        }
        // if any of bigTableCandidates is from multi-sourced, bigTableCandidates should
        // only contain multi-sourced because multi-sourced cannot be hashed or direct readable
        bigTableCandidates = multiInsertBigTableCheck(joinOp, bigTableCandidates);
        Configuration conf = context.getConf();
        // If sizes of at least n-1 tables in a n-way join is known, and their sum is smaller than
        // the threshold size, convert the join into map-join and don't create a conditional task
        boolean convertJoinMapJoin = HiveConf.getBoolVar(conf, HiveConf.ConfVars.HIVECONVERTJOINNOCONDITIONALTASK);
        int bigTablePosition = -1;
        if (convertJoinMapJoin) {
            // This is the threshold that the user has specified to fit in mapjoin
            long mapJoinSize = HiveConf.getLongVar(conf, HiveConf.ConfVars.HIVECONVERTJOINNOCONDITIONALTASKTHRESHOLD);
            Long bigTableSize = null;
            Set<String> aliases = aliasToWork.keySet();
            for (int tablePosition : bigTableCandidates) {
                Operator<?> parent = joinOp.getParentOperators().get(tablePosition);
                Set<String> participants = GenMapRedUtils.findAliases(currWork, parent);
                long sumOfOthers = Utilities.sumOfExcept(aliasToSize, aliases, participants);
                if (sumOfOthers < 0 || sumOfOthers > mapJoinSize) {
                    // some small alias is not known or too big
                    continue;
                }
                if (bigTableSize == null && bigTablePosition >= 0 && tablePosition < bigTablePosition) {
                    // prefer right most alias
                    continue;
                }
                long aliasSize = Utilities.sumOf(aliasToSize, participants);
                if (bigTableSize == null || bigTableSize < 0 || (aliasSize >= 0 && aliasSize >= bigTableSize)) {
                    bigTablePosition = tablePosition;
                    bigTableSize = aliasSize;
                }
            }
        }
        currWork.setLeftInputJoin(joinOp.getConf().isLeftInputJoin());
        currWork.setBaseSrc(joinOp.getConf().getBaseSrc());
        currWork.setMapAliases(joinOp.getConf().getMapAliases());
        if (bigTablePosition >= 0) {
            // create map join task and set big table as bigTablePosition
            MapRedTask newTask = convertTaskToMapJoinTask(currTask.getWork(), bigTablePosition);
            newTask.setTaskTag(Task.MAPJOIN_ONLY_NOBACKUP);
            newTask.setFetchSource(currTask.isFetchSource());
            replaceTask(currTask, newTask);
            // joined with multiple small tables on different keys
            if ((newTask.getChildTasks() != null) && (newTask.getChildTasks().size() == 1)) {
                mergeMapJoinTaskIntoItsChildMapRedTask(newTask, conf);
            }
            return newTask;
        }
        long ThresholdOfSmallTblSizeSum = HiveConf.getLongVar(conf, HiveConf.ConfVars.HIVESMALLTABLESFILESIZE);
        for (int pos = 0; pos < joinOp.getNumParent(); pos++) {
            // this table cannot be big table
            if (!bigTableCandidates.contains(pos)) {
                continue;
            }
            Operator<?> startOp = joinOp.getParentOperators().get(pos);
            Set<String> aliases = GenMapRedUtils.findAliases(currWork, startOp);
            long aliasKnownSize = Utilities.sumOf(aliasToSize, aliases);
            if (cannotConvert(aliasKnownSize, aliasTotalKnownInputSize, ThresholdOfSmallTblSizeSum)) {
                continue;
            }
            MapredWork newWork = SerializationUtilities.clonePlan(currTask.getWork());
            // create map join task and set big table as i
            MapRedTask newTask = convertTaskToMapJoinTask(newWork, pos);
            // add into conditional task
            listWorks.add(newTask.getWork());
            listTasks.add(newTask);
            newTask.setTaskTag(Task.CONVERTED_MAPJOIN);
            newTask.setFetchSource(currTask.isFetchSource());
            // set up backup task
            newTask.setBackupTask(currTask);
            newTask.setBackupChildrenTasks(currTask.getChildTasks());
            // put the mapping task to aliases
            taskToAliases.put(newTask, aliases);
        }
    } catch (Exception e) {
        throw new SemanticException("Generate Map Join Task Error: " + e.getMessage(), e);
    }
    if (listTasks.isEmpty()) {
        return currTask;
    }
    // insert current common join task to conditional task
    listWorks.add(currTask.getWork());
    listTasks.add(currTask);
    // clear JoinTree and OP Parse Context
    currWork.setLeftInputJoin(false);
    currWork.setBaseSrc(null);
    currWork.setMapAliases(null);
    // create conditional task and insert conditional task into task tree
    ConditionalWork cndWork = new ConditionalWork(listWorks);
    ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork);
    cndTsk.setListTasks(listTasks);
    // set resolver and resolver context
    cndTsk.setResolver(new ConditionalResolverCommonJoin());
    ConditionalResolverCommonJoinCtx resolverCtx = new ConditionalResolverCommonJoinCtx();
    resolverCtx.setPathToAliases(pathToAliases);
    resolverCtx.setAliasToKnownSize(aliasToSize);
    resolverCtx.setTaskToAliases(taskToAliases);
    resolverCtx.setCommonJoinTask(currTask);
    resolverCtx.setLocalTmpDir(context.getLocalScratchDir(false));
    resolverCtx.setHdfsTmpDir(context.getMRScratchDir());
    cndTsk.setResolverCtx(resolverCtx);
    // replace the current task with the new generated conditional task
    replaceTaskWithConditionalTask(currTask, cndTsk);
    return cndTsk;
}
Also used : JoinOperator(org.apache.hadoop.hive.ql.exec.JoinOperator) LateralViewForwardOperator(org.apache.hadoop.hive.ql.exec.LateralViewForwardOperator) FileSinkOperator(org.apache.hadoop.hive.ql.exec.FileSinkOperator) JoinOperator(org.apache.hadoop.hive.ql.exec.JoinOperator) TableScanOperator(org.apache.hadoop.hive.ql.exec.TableScanOperator) Operator(org.apache.hadoop.hive.ql.exec.Operator) Serializable(java.io.Serializable) ConditionalTask(org.apache.hadoop.hive.ql.exec.ConditionalTask) Task(org.apache.hadoop.hive.ql.exec.Task) MapRedTask(org.apache.hadoop.hive.ql.exec.mr.MapRedTask) HashSet(java.util.HashSet) Set(java.util.Set) Configuration(org.apache.hadoop.conf.Configuration) ArrayList(java.util.ArrayList) ConditionalWork(org.apache.hadoop.hive.ql.plan.ConditionalWork) LinkedHashMap(java.util.LinkedHashMap) MapRedTask(org.apache.hadoop.hive.ql.exec.mr.MapRedTask) ConditionalResolverCommonJoinCtx(org.apache.hadoop.hive.ql.plan.ConditionalResolverCommonJoin.ConditionalResolverCommonJoinCtx) MapredWork(org.apache.hadoop.hive.ql.plan.MapredWork) ConditionalTask(org.apache.hadoop.hive.ql.exec.ConditionalTask) ArrayList(java.util.ArrayList) List(java.util.List) ConditionalResolverCommonJoin(org.apache.hadoop.hive.ql.plan.ConditionalResolverCommonJoin) SemanticException(org.apache.hadoop.hive.ql.parse.SemanticException) Path(org.apache.hadoop.fs.Path) SemanticException(org.apache.hadoop.hive.ql.parse.SemanticException) MapWork(org.apache.hadoop.hive.ql.plan.MapWork) OperatorDesc(org.apache.hadoop.hive.ql.plan.OperatorDesc) JoinDesc(org.apache.hadoop.hive.ql.plan.JoinDesc)

Aggregations

Serializable (java.io.Serializable)2 ArrayList (java.util.ArrayList)2 LinkedHashMap (java.util.LinkedHashMap)2 List (java.util.List)2 Set (java.util.Set)2 Configuration (org.apache.hadoop.conf.Configuration)2 Path (org.apache.hadoop.fs.Path)2 ConditionalTask (org.apache.hadoop.hive.ql.exec.ConditionalTask)2 Task (org.apache.hadoop.hive.ql.exec.Task)2 MapRedTask (org.apache.hadoop.hive.ql.exec.mr.MapRedTask)2 SemanticException (org.apache.hadoop.hive.ql.parse.SemanticException)2 ConditionalResolverCommonJoin (org.apache.hadoop.hive.ql.plan.ConditionalResolverCommonJoin)2 ConditionalResolverCommonJoinCtx (org.apache.hadoop.hive.ql.plan.ConditionalResolverCommonJoin.ConditionalResolverCommonJoinCtx)2 ConditionalWork (org.apache.hadoop.hive.ql.plan.ConditionalWork)2 MapWork (org.apache.hadoop.hive.ql.plan.MapWork)2 MapredWork (org.apache.hadoop.hive.ql.plan.MapredWork)2 HashMap (java.util.HashMap)1 HashSet (java.util.HashSet)1 FileSinkOperator (org.apache.hadoop.hive.ql.exec.FileSinkOperator)1 JoinOperator (org.apache.hadoop.hive.ql.exec.JoinOperator)1