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

Example 1 with VectorColumnSourceMapping

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

the class Vectorizer method canSpecializeMapJoin.

private boolean canSpecializeMapJoin(Operator<? extends OperatorDesc> op, MapJoinDesc desc, boolean isTezOrSpark, VectorizationContext vContext, VectorMapJoinInfo vectorMapJoinInfo) throws HiveException {
    Preconditions.checkState(op instanceof MapJoinOperator);
    // Allocate a VectorReduceSinkDesc initially with implementation type NONE so EXPLAIN
    // can report this operator was vectorized, but not native.  And, the conditions.
    VectorMapJoinDesc vectorDesc = new VectorMapJoinDesc();
    desc.setVectorDesc(vectorDesc);
    boolean isVectorizationMapJoinNativeEnabled = HiveConf.getBoolVar(hiveConf, HiveConf.ConfVars.HIVE_VECTORIZATION_MAPJOIN_NATIVE_ENABLED);
    String engine = HiveConf.getVar(hiveConf, HiveConf.ConfVars.HIVE_EXECUTION_ENGINE);
    boolean oneMapJoinCondition = (desc.getConds().length == 1);
    boolean hasNullSafes = onExpressionHasNullSafes(desc);
    byte posBigTable = (byte) desc.getPosBigTable();
    // Since we want to display all the met and not met conditions in EXPLAIN, we determine all
    // information first....
    List<ExprNodeDesc> keyDesc = desc.getKeys().get(posBigTable);
    VectorExpression[] allBigTableKeyExpressions = vContext.getVectorExpressions(keyDesc);
    final int allBigTableKeyExpressionsLength = allBigTableKeyExpressions.length;
    // Assume.
    boolean supportsKeyTypes = true;
    HashSet<String> notSupportedKeyTypes = new HashSet<String>();
    // Since a key expression can be a calculation and the key will go into a scratch column,
    // we need the mapping and type information.
    int[] bigTableKeyColumnMap = new int[allBigTableKeyExpressionsLength];
    String[] bigTableKeyColumnNames = new String[allBigTableKeyExpressionsLength];
    TypeInfo[] bigTableKeyTypeInfos = new TypeInfo[allBigTableKeyExpressionsLength];
    ArrayList<VectorExpression> bigTableKeyExpressionsList = new ArrayList<VectorExpression>();
    VectorExpression[] bigTableKeyExpressions;
    for (int i = 0; i < allBigTableKeyExpressionsLength; i++) {
        VectorExpression ve = allBigTableKeyExpressions[i];
        if (!IdentityExpression.isColumnOnly(ve)) {
            bigTableKeyExpressionsList.add(ve);
        }
        bigTableKeyColumnMap[i] = ve.getOutputColumn();
        ExprNodeDesc exprNode = keyDesc.get(i);
        bigTableKeyColumnNames[i] = exprNode.toString();
        TypeInfo typeInfo = exprNode.getTypeInfo();
        // same check used in HashTableLoader.
        if (!MapJoinKey.isSupportedField(typeInfo)) {
            supportsKeyTypes = false;
            Category category = typeInfo.getCategory();
            notSupportedKeyTypes.add((category != Category.PRIMITIVE ? category.toString() : ((PrimitiveTypeInfo) typeInfo).getPrimitiveCategory().toString()));
        }
        bigTableKeyTypeInfos[i] = typeInfo;
    }
    if (bigTableKeyExpressionsList.size() == 0) {
        bigTableKeyExpressions = null;
    } else {
        bigTableKeyExpressions = bigTableKeyExpressionsList.toArray(new VectorExpression[0]);
    }
    List<ExprNodeDesc> bigTableExprs = desc.getExprs().get(posBigTable);
    VectorExpression[] allBigTableValueExpressions = vContext.getVectorExpressions(bigTableExprs);
    boolean isFastHashTableEnabled = HiveConf.getBoolVar(hiveConf, HiveConf.ConfVars.HIVE_VECTORIZATION_MAPJOIN_NATIVE_FAST_HASHTABLE_ENABLED);
    // Especially since LLAP is prone to turn it off in the MapJoinDesc in later
    // physical optimizer stages...
    boolean isHybridHashJoin = desc.isHybridHashJoin();
    /*
     * Populate vectorMapJoininfo.
     */
    /*
     * Similarly, we need a mapping since a value expression can be a calculation and the value
     * will go into a scratch column.
     */
    int[] bigTableValueColumnMap = new int[allBigTableValueExpressions.length];
    String[] bigTableValueColumnNames = new String[allBigTableValueExpressions.length];
    TypeInfo[] bigTableValueTypeInfos = new TypeInfo[allBigTableValueExpressions.length];
    ArrayList<VectorExpression> bigTableValueExpressionsList = new ArrayList<VectorExpression>();
    VectorExpression[] bigTableValueExpressions;
    for (int i = 0; i < bigTableValueColumnMap.length; i++) {
        VectorExpression ve = allBigTableValueExpressions[i];
        if (!IdentityExpression.isColumnOnly(ve)) {
            bigTableValueExpressionsList.add(ve);
        }
        bigTableValueColumnMap[i] = ve.getOutputColumn();
        ExprNodeDesc exprNode = bigTableExprs.get(i);
        bigTableValueColumnNames[i] = exprNode.toString();
        bigTableValueTypeInfos[i] = exprNode.getTypeInfo();
    }
    if (bigTableValueExpressionsList.size() == 0) {
        bigTableValueExpressions = null;
    } else {
        bigTableValueExpressions = bigTableValueExpressionsList.toArray(new VectorExpression[0]);
    }
    vectorMapJoinInfo.setBigTableKeyColumnMap(bigTableKeyColumnMap);
    vectorMapJoinInfo.setBigTableKeyColumnNames(bigTableKeyColumnNames);
    vectorMapJoinInfo.setBigTableKeyTypeInfos(bigTableKeyTypeInfos);
    vectorMapJoinInfo.setBigTableKeyExpressions(bigTableKeyExpressions);
    vectorMapJoinInfo.setBigTableValueColumnMap(bigTableValueColumnMap);
    vectorMapJoinInfo.setBigTableValueColumnNames(bigTableValueColumnNames);
    vectorMapJoinInfo.setBigTableValueTypeInfos(bigTableValueTypeInfos);
    vectorMapJoinInfo.setBigTableValueExpressions(bigTableValueExpressions);
    /*
     * Small table information.
     */
    VectorColumnOutputMapping bigTableRetainedMapping = new VectorColumnOutputMapping("Big Table Retained Mapping");
    VectorColumnOutputMapping bigTableOuterKeyMapping = new VectorColumnOutputMapping("Big Table Outer Key Mapping");
    // The order of the fields in the LazyBinary small table value must be used, so
    // we use the source ordering flavor for the mapping.
    VectorColumnSourceMapping smallTableMapping = new VectorColumnSourceMapping("Small Table Mapping");
    Byte[] order = desc.getTagOrder();
    Byte posSingleVectorMapJoinSmallTable = (order[0] == posBigTable ? order[1] : order[0]);
    boolean isOuterJoin = !desc.getNoOuterJoin();
    /*
     * Gather up big and small table output result information from the MapJoinDesc.
     */
    List<Integer> bigTableRetainList = desc.getRetainList().get(posBigTable);
    int bigTableRetainSize = bigTableRetainList.size();
    int[] smallTableIndices;
    int smallTableIndicesSize;
    List<ExprNodeDesc> smallTableExprs = desc.getExprs().get(posSingleVectorMapJoinSmallTable);
    if (desc.getValueIndices() != null && desc.getValueIndices().get(posSingleVectorMapJoinSmallTable) != null) {
        smallTableIndices = desc.getValueIndices().get(posSingleVectorMapJoinSmallTable);
        smallTableIndicesSize = smallTableIndices.length;
    } else {
        smallTableIndices = null;
        smallTableIndicesSize = 0;
    }
    List<Integer> smallTableRetainList = desc.getRetainList().get(posSingleVectorMapJoinSmallTable);
    int smallTableRetainSize = smallTableRetainList.size();
    int smallTableResultSize = 0;
    if (smallTableIndicesSize > 0) {
        smallTableResultSize = smallTableIndicesSize;
    } else if (smallTableRetainSize > 0) {
        smallTableResultSize = smallTableRetainSize;
    }
    /*
     * Determine the big table retained mapping first so we can optimize out (with
     * projection) copying inner join big table keys in the subsequent small table results section.
     */
    // We use a mapping object here so we can build the projection in any order and
    // get the ordered by 0 to n-1 output columns at the end.
    //
    // Also, to avoid copying a big table key into the small table result area for inner joins,
    // we reference it with the projection so there can be duplicate output columns
    // in the projection.
    VectorColumnSourceMapping projectionMapping = new VectorColumnSourceMapping("Projection Mapping");
    int nextOutputColumn = (order[0] == posBigTable ? 0 : smallTableResultSize);
    for (int i = 0; i < bigTableRetainSize; i++) {
        // Since bigTableValueExpressions may do a calculation and produce a scratch column, we
        // need to map to the right batch column.
        int retainColumn = bigTableRetainList.get(i);
        int batchColumnIndex = bigTableValueColumnMap[retainColumn];
        TypeInfo typeInfo = bigTableValueTypeInfos[i];
        // With this map we project the big table batch to make it look like an output batch.
        projectionMapping.add(nextOutputColumn, batchColumnIndex, typeInfo);
        // Collect columns we copy from the big table batch to the overflow batch.
        if (!bigTableRetainedMapping.containsOutputColumn(batchColumnIndex)) {
            // Tolerate repeated use of a big table column.
            bigTableRetainedMapping.add(batchColumnIndex, batchColumnIndex, typeInfo);
        }
        nextOutputColumn++;
    }
    /*
     * Now determine the small table results.
     */
    boolean smallTableExprVectorizes = true;
    int firstSmallTableOutputColumn;
    firstSmallTableOutputColumn = (order[0] == posBigTable ? bigTableRetainSize : 0);
    int smallTableOutputCount = 0;
    nextOutputColumn = firstSmallTableOutputColumn;
    // Small table indices has more information (i.e. keys) than retain, so use it if it exists...
    String[] bigTableRetainedNames;
    if (smallTableIndicesSize > 0) {
        smallTableOutputCount = smallTableIndicesSize;
        bigTableRetainedNames = new String[smallTableOutputCount];
        for (int i = 0; i < smallTableIndicesSize; i++) {
            if (smallTableIndices[i] >= 0) {
                // Zero and above numbers indicate a big table key is needed for
                // small table result "area".
                int keyIndex = smallTableIndices[i];
                // Since bigTableKeyExpressions may do a calculation and produce a scratch column, we
                // need to map the right column.
                int batchKeyColumn = bigTableKeyColumnMap[keyIndex];
                bigTableRetainedNames[i] = bigTableKeyColumnNames[keyIndex];
                TypeInfo typeInfo = bigTableKeyTypeInfos[keyIndex];
                if (!isOuterJoin) {
                    // Optimize inner join keys of small table results.
                    // Project the big table key into the small table result "area".
                    projectionMapping.add(nextOutputColumn, batchKeyColumn, typeInfo);
                    if (!bigTableRetainedMapping.containsOutputColumn(batchKeyColumn)) {
                        // If necessary, copy the big table key into the overflow batch's small table
                        // result "area".
                        bigTableRetainedMapping.add(batchKeyColumn, batchKeyColumn, typeInfo);
                    }
                } else {
                    // For outer joins, since the small table key can be null when there is no match,
                    // we must have a physical (scratch) column for those keys.  We cannot use the
                    // projection optimization used by inner joins above.
                    int scratchColumn = vContext.allocateScratchColumn(typeInfo);
                    projectionMapping.add(nextOutputColumn, scratchColumn, typeInfo);
                    bigTableRetainedMapping.add(batchKeyColumn, scratchColumn, typeInfo);
                    bigTableOuterKeyMapping.add(batchKeyColumn, scratchColumn, typeInfo);
                }
            } else {
                // Negative numbers indicate a column to be (deserialize) read from the small table's
                // LazyBinary value row.
                int smallTableValueIndex = -smallTableIndices[i] - 1;
                ExprNodeDesc smallTableExprNode = smallTableExprs.get(i);
                if (!validateExprNodeDesc(smallTableExprNode, "Small Table")) {
                    clearNotVectorizedReason();
                    smallTableExprVectorizes = false;
                }
                bigTableRetainedNames[i] = smallTableExprNode.toString();
                TypeInfo typeInfo = smallTableExprNode.getTypeInfo();
                // Make a new big table scratch column for the small table value.
                int scratchColumn = vContext.allocateScratchColumn(typeInfo);
                projectionMapping.add(nextOutputColumn, scratchColumn, typeInfo);
                smallTableMapping.add(smallTableValueIndex, scratchColumn, typeInfo);
            }
            nextOutputColumn++;
        }
    } else if (smallTableRetainSize > 0) {
        smallTableOutputCount = smallTableRetainSize;
        bigTableRetainedNames = new String[smallTableOutputCount];
        for (int i = 0; i < smallTableRetainSize; i++) {
            int smallTableValueIndex = smallTableRetainList.get(i);
            ExprNodeDesc smallTableExprNode = smallTableExprs.get(i);
            if (!validateExprNodeDesc(smallTableExprNode, "Small Table")) {
                clearNotVectorizedReason();
                smallTableExprVectorizes = false;
            }
            bigTableRetainedNames[i] = smallTableExprNode.toString();
            // Make a new big table scratch column for the small table value.
            TypeInfo typeInfo = smallTableExprNode.getTypeInfo();
            int scratchColumn = vContext.allocateScratchColumn(typeInfo);
            projectionMapping.add(nextOutputColumn, scratchColumn, typeInfo);
            smallTableMapping.add(smallTableValueIndex, scratchColumn, typeInfo);
            nextOutputColumn++;
        }
    } else {
        bigTableRetainedNames = new String[0];
    }
    boolean useOptimizedTable = HiveConf.getBoolVar(hiveConf, HiveConf.ConfVars.HIVEMAPJOINUSEOPTIMIZEDTABLE);
    // Remember the condition variables for EXPLAIN regardless of whether we specialize or not.
    vectorDesc.setUseOptimizedTable(useOptimizedTable);
    vectorDesc.setIsVectorizationMapJoinNativeEnabled(isVectorizationMapJoinNativeEnabled);
    vectorDesc.setEngine(engine);
    vectorDesc.setOneMapJoinCondition(oneMapJoinCondition);
    vectorDesc.setHasNullSafes(hasNullSafes);
    vectorDesc.setSmallTableExprVectorizes(smallTableExprVectorizes);
    vectorDesc.setIsFastHashTableEnabled(isFastHashTableEnabled);
    vectorDesc.setIsHybridHashJoin(isHybridHashJoin);
    vectorDesc.setSupportsKeyTypes(supportsKeyTypes);
    if (!supportsKeyTypes) {
        vectorDesc.setNotSupportedKeyTypes(new ArrayList(notSupportedKeyTypes));
    }
    // Check common conditions for both Optimized and Fast Hash Tables.
    // Assume.
    boolean result = true;
    if (!useOptimizedTable || !isVectorizationMapJoinNativeEnabled || !isTezOrSpark || !oneMapJoinCondition || hasNullSafes || !smallTableExprVectorizes) {
        result = false;
    }
    if (!isFastHashTableEnabled) {
        // Check optimized-only hash table restrictions.
        if (!supportsKeyTypes) {
            result = false;
        }
    } else {
        if (isHybridHashJoin) {
            result = false;
        }
    }
    // Convert dynamic arrays and maps to simple arrays.
    bigTableRetainedMapping.finalize();
    bigTableOuterKeyMapping.finalize();
    smallTableMapping.finalize();
    vectorMapJoinInfo.setBigTableRetainedMapping(bigTableRetainedMapping);
    vectorMapJoinInfo.setBigTableOuterKeyMapping(bigTableOuterKeyMapping);
    vectorMapJoinInfo.setSmallTableMapping(smallTableMapping);
    projectionMapping.finalize();
    // Verify we added an entry for each output.
    assert projectionMapping.isSourceSequenceGood();
    vectorMapJoinInfo.setProjectionMapping(projectionMapping);
    return result;
}