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

Example 6 with ColStatistics

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

the class FilterSelectivityEstimator method getMaxNulls.

/**
   * Given a RexCall & TableScan find max no of nulls. Currently it picks the
   * col with max no of nulls.
   * 
   * TODO: improve this
   * 
   * @param call
   * @param t
   * @return
   */
private long getMaxNulls(RexCall call, HiveTableScan t) {
    long tmpNoNulls = 0;
    long maxNoNulls = 0;
    Set<Integer> iRefSet = HiveCalciteUtil.getInputRefs(call);
    List<ColStatistics> colStats = t.getColStat(new ArrayList<Integer>(iRefSet));
    for (ColStatistics cs : colStats) {
        tmpNoNulls = cs.getNumNulls();
        if (tmpNoNulls > maxNoNulls) {
            maxNoNulls = tmpNoNulls;
        }
    }
    return maxNoNulls;
}

Example 7 with ColStatistics

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

the class StatsUtils method inferForeignKey.

/**
   * Infer foreign key relationship from given column statistics.
   * @param csPK - column statistics of primary key
   * @param csFK - column statistics of potential foreign key
   * @return
   */
public static boolean inferForeignKey(ColStatistics csPK, ColStatistics csFK) {
    if (csPK != null && csFK != null) {
        if (csPK.isPrimaryKey()) {
            if (csPK.getRange() != null && csFK.getRange() != null) {
                ColStatistics.Range pkRange = csPK.getRange();
                ColStatistics.Range fkRange = csFK.getRange();
                return isWithin(fkRange, pkRange);
            }
        }
    }
    return false;
}

Example 8 with ColStatistics

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

the class ConvertJoinMapJoin method checkNumberOfEntriesForHashTable.

/* Returns true if it passes the test, false otherwise. */
private boolean checkNumberOfEntriesForHashTable(JoinOperator joinOp, int position, OptimizeTezProcContext context) {
    long max = HiveConf.getLongVar(context.parseContext.getConf(), HiveConf.ConfVars.HIVECONVERTJOINMAXENTRIESHASHTABLE);
    if (max < 1) {
        // Max is disabled, we can safely return true
        return true;
    }
    // Calculate number of different entries and evaluate
    ReduceSinkOperator rsOp = (ReduceSinkOperator) joinOp.getParentOperators().get(position);
    List<String> keys = StatsUtils.getQualifedReducerKeyNames(rsOp.getConf().getOutputKeyColumnNames());
    Statistics inputStats = rsOp.getStatistics();
    List<ColStatistics> columnStats = new ArrayList<>();
    for (String key : keys) {
        ColStatistics cs = inputStats.getColumnStatisticsFromColName(key);
        if (cs == null) {
            LOG.debug("Couldn't get statistics for: {}", key);
            return true;
        }
        columnStats.add(cs);
    }
    long numRows = inputStats.getNumRows();
    long estimation = estimateNDV(numRows, columnStats);
    LOG.debug("Estimated NDV for input {}: {}; Max NDV for MapJoin conversion: {}", position, estimation, max);
    if (estimation > max) {
        // Estimation larger than max
        LOG.debug("Number of different entries for HashTable is greater than the max; " + "we do not converting to MapJoin");
        return false;
    }
    // We can proceed with the conversion
    return true;
}

Example 9 with ColStatistics

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

the class ConvertJoinMapJoin method estimateNDV.

private static long estimateNDV(long numRows, List<ColStatistics> columnStats) {
    // If there is a single column, return the number of distinct values
    if (columnStats.size() == 1) {
        return columnStats.get(0).getCountDistint();
    }
    // The expected number of distinct values when choosing p values
    // with replacement from n integers is n . (1 - ((n - 1) / n) ^ p).
    //
    // If we have several uniformly distributed attributes A1 ... Am
    // with N1 ... Nm distinct values, they behave as one uniformly
    // distributed attribute with N1 * ... * Nm distinct values.
    long n = 1L;
    for (ColStatistics cs : columnStats) {
        final long ndv = cs.getCountDistint();
        if (ndv > 1) {
            n = StatsUtils.safeMult(n, ndv);
        }
    }
    final double nn = (double) n;
    final double a = (nn - 1d) / nn;
    if (a == 1d) {
        // A under-flows if nn is large.
        return numRows;
    }
    final double v = nn * (1d - Math.pow(a, numRows));
    // to go a few % over.
    return Math.min(Math.round(v), numRows);
}

Example 10 with ColStatistics

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

the class ReduceSinkMapJoinProc method processReduceSinkToHashJoin.

public static Object processReduceSinkToHashJoin(ReduceSinkOperator parentRS, MapJoinOperator mapJoinOp, GenTezProcContext context) throws SemanticException {
    // remove the tag for in-memory side of mapjoin
    parentRS.getConf().setSkipTag(true);
    parentRS.setSkipTag(true);
    // Mark this small table as being processed
    if (mapJoinOp.getConf().isDynamicPartitionHashJoin()) {
        context.mapJoinToUnprocessedSmallTableReduceSinks.get(mapJoinOp).remove(parentRS);
    }
    List<BaseWork> mapJoinWork = null;
    /*
     *  if there was a pre-existing work generated for the big-table mapjoin side,
     *  we need to hook the work generated for the RS (associated with the RS-MJ pattern)
     *  with the pre-existing work.
     *
     *  Otherwise, we need to associate that the mapjoin op
     *  to be linked to the RS work (associated with the RS-MJ pattern).
     *
     */
    mapJoinWork = context.mapJoinWorkMap.get(mapJoinOp);
    BaseWork parentWork = getMapJoinParentWork(context, parentRS);
    // set the link between mapjoin and parent vertex
    int pos = context.mapJoinParentMap.get(mapJoinOp).indexOf(parentRS);
    if (pos == -1) {
        throw new SemanticException("Cannot find position of parent in mapjoin");
    }
    MapJoinDesc joinConf = mapJoinOp.getConf();
    long keyCount = Long.MAX_VALUE, rowCount = Long.MAX_VALUE, bucketCount = 1;
    long tableSize = Long.MAX_VALUE;
    Statistics stats = parentRS.getStatistics();
    if (stats != null) {
        keyCount = rowCount = stats.getNumRows();
        if (keyCount <= 0) {
            keyCount = rowCount = Long.MAX_VALUE;
        }
        tableSize = stats.getDataSize();
        ArrayList<String> keyCols = parentRS.getConf().getOutputKeyColumnNames();
        if (keyCols != null && !keyCols.isEmpty()) {
            // See if we can arrive at a smaller number using distinct stats from key columns.
            long maxKeyCount = 1;
            String prefix = Utilities.ReduceField.KEY.toString();
            for (String keyCol : keyCols) {
                ExprNodeDesc realCol = parentRS.getColumnExprMap().get(prefix + "." + keyCol);
                ColStatistics cs = StatsUtils.getColStatisticsFromExpression(context.conf, stats, realCol);
                if (cs == null || cs.getCountDistint() <= 0) {
                    maxKeyCount = Long.MAX_VALUE;
                    break;
                }
                maxKeyCount *= cs.getCountDistint();
                if (maxKeyCount >= keyCount) {
                    break;
                }
            }
            keyCount = Math.min(maxKeyCount, keyCount);
        }
        if (joinConf.isBucketMapJoin()) {
            OpTraits opTraits = mapJoinOp.getOpTraits();
            bucketCount = (opTraits == null) ? -1 : opTraits.getNumBuckets();
            if (bucketCount > 0) {
                // We cannot obtain a better estimate without CustomPartitionVertex providing it
                // to us somehow; in which case using statistics would be completely unnecessary.
                keyCount /= bucketCount;
                tableSize /= bucketCount;
            }
        } else if (joinConf.isDynamicPartitionHashJoin()) {
            // For dynamic partitioned hash join, assuming table is split evenly among the reduce tasks.
            bucketCount = parentRS.getConf().getNumReducers();
            keyCount /= bucketCount;
            tableSize /= bucketCount;
        }
    }
    if (keyCount == 0) {
        keyCount = 1;
    }
    if (tableSize == 0) {
        tableSize = 1;
    }
    LOG.info("Mapjoin " + mapJoinOp + "(bucket map join = )" + joinConf.isBucketMapJoin() + ", pos: " + pos + " --> " + parentWork.getName() + " (" + keyCount + " keys estimated from " + rowCount + " rows, " + bucketCount + " buckets)");
    joinConf.getParentToInput().put(pos, parentWork.getName());
    if (keyCount != Long.MAX_VALUE) {
        joinConf.getParentKeyCounts().put(pos, keyCount);
    }
    joinConf.getParentDataSizes().put(pos, tableSize);
    int numBuckets = -1;
    EdgeType edgeType = EdgeType.BROADCAST_EDGE;
    if (joinConf.isBucketMapJoin()) {
        numBuckets = (Integer) joinConf.getBigTableBucketNumMapping().values().toArray()[0];
        /*
       * Here, we can be in one of 4 states.
       *
       * 1. If map join work is null implies that we have not yet traversed the big table side. We
       * just need to see if we can find a reduce sink operator in the big table side. This would
       * imply a reduce side operation.
       *
       * 2. If we don't find a reducesink in 1 it has to be the case that it is a map side operation.
       *
       * 3. If we have already created a work item for the big table side, we need to see if we can
       * find a table scan operator in the big table side. This would imply a map side operation.
       *
       * 4. If we don't find a table scan operator, it has to be a reduce side operation.
       */
        if (mapJoinWork == null) {
            Operator<?> rootOp = OperatorUtils.findSingleOperatorUpstreamJoinAccounted(mapJoinOp.getParentOperators().get(joinConf.getPosBigTable()), ReduceSinkOperator.class);
            if (rootOp == null) {
                // likely we found a table scan operator
                edgeType = EdgeType.CUSTOM_EDGE;
            } else {
                // we have found a reduce sink
                edgeType = EdgeType.CUSTOM_SIMPLE_EDGE;
            }
        } else {
            Operator<?> rootOp = OperatorUtils.findSingleOperatorUpstreamJoinAccounted(mapJoinOp.getParentOperators().get(joinConf.getPosBigTable()), TableScanOperator.class);
            if (rootOp != null) {
                // likely we found a table scan operator
                edgeType = EdgeType.CUSTOM_EDGE;
            } else {
                // we have found a reduce sink
                edgeType = EdgeType.CUSTOM_SIMPLE_EDGE;
            }
        }
    } else if (mapJoinOp.getConf().isDynamicPartitionHashJoin()) {
        edgeType = EdgeType.CUSTOM_SIMPLE_EDGE;
    }
    if (edgeType == EdgeType.CUSTOM_EDGE) {
        // disable auto parallelism for bucket map joins
        parentRS.getConf().setReducerTraits(EnumSet.of(FIXED));
    }
    TezEdgeProperty edgeProp = new TezEdgeProperty(null, edgeType, numBuckets);
    if (mapJoinWork != null) {
        for (BaseWork myWork : mapJoinWork) {
            // link the work with the work associated with the reduce sink that triggered this rule
            TezWork tezWork = context.currentTask.getWork();
            LOG.debug("connecting " + parentWork.getName() + " with " + myWork.getName());
            tezWork.connect(parentWork, myWork, edgeProp);
            if (edgeType == EdgeType.CUSTOM_EDGE) {
                tezWork.setVertexType(myWork, VertexType.INITIALIZED_EDGES);
            }
            ReduceSinkOperator r = null;
            if (context.connectedReduceSinks.contains(parentRS)) {
                LOG.debug("Cloning reduce sink for multi-child broadcast edge");
                // we've already set this one up. Need to clone for the next work.
                r = (ReduceSinkOperator) OperatorFactory.getAndMakeChild(parentRS.getCompilationOpContext(), (ReduceSinkDesc) parentRS.getConf().clone(), new RowSchema(parentRS.getSchema()), parentRS.getParentOperators());
                context.clonedReduceSinks.add(r);
            } else {
                r = parentRS;
            }
            // remember the output name of the reduce sink
            r.getConf().setOutputName(myWork.getName());
            context.connectedReduceSinks.add(r);
        }
    }
    // remember in case we need to connect additional work later
    Map<BaseWork, TezEdgeProperty> linkWorkMap = null;
    if (context.linkOpWithWorkMap.containsKey(mapJoinOp)) {
        linkWorkMap = context.linkOpWithWorkMap.get(mapJoinOp);
    } else {
        linkWorkMap = new HashMap<BaseWork, TezEdgeProperty>();
    }
    linkWorkMap.put(parentWork, edgeProp);
    context.linkOpWithWorkMap.put(mapJoinOp, linkWorkMap);
    List<ReduceSinkOperator> reduceSinks = context.linkWorkWithReduceSinkMap.get(parentWork);
    if (reduceSinks == null) {
        reduceSinks = new ArrayList<ReduceSinkOperator>();
    }
    reduceSinks.add(parentRS);
    context.linkWorkWithReduceSinkMap.put(parentWork, reduceSinks);
    // create the dummy operators
    List<Operator<?>> dummyOperators = new ArrayList<Operator<?>>();
    // create an new operator: HashTableDummyOperator, which share the table desc
    HashTableDummyDesc desc = new HashTableDummyDesc();
    @SuppressWarnings("unchecked") HashTableDummyOperator dummyOp = (HashTableDummyOperator) OperatorFactory.get(parentRS.getCompilationOpContext(), desc);
    TableDesc tbl;
    // need to create the correct table descriptor for key/value
    RowSchema rowSchema = parentRS.getParentOperators().get(0).getSchema();
    tbl = PlanUtils.getReduceValueTableDesc(PlanUtils.getFieldSchemasFromRowSchema(rowSchema, ""));
    dummyOp.getConf().setTbl(tbl);
    Map<Byte, List<ExprNodeDesc>> keyExprMap = mapJoinOp.getConf().getKeys();
    List<ExprNodeDesc> keyCols = keyExprMap.get(Byte.valueOf((byte) 0));
    StringBuilder keyOrder = new StringBuilder();
    StringBuilder keyNullOrder = new StringBuilder();
    for (ExprNodeDesc k : keyCols) {
        keyOrder.append("+");
        keyNullOrder.append("a");
    }
    TableDesc keyTableDesc = PlanUtils.getReduceKeyTableDesc(PlanUtils.getFieldSchemasFromColumnList(keyCols, "mapjoinkey"), keyOrder.toString(), keyNullOrder.toString());
    mapJoinOp.getConf().setKeyTableDesc(keyTableDesc);
    // let the dummy op be the parent of mapjoin op
    mapJoinOp.replaceParent(parentRS, dummyOp);
    List<Operator<? extends OperatorDesc>> dummyChildren = new ArrayList<Operator<? extends OperatorDesc>>();
    dummyChildren.add(mapJoinOp);
    dummyOp.setChildOperators(dummyChildren);
    dummyOperators.add(dummyOp);
    // cut the operator tree so as to not retain connections from the parent RS downstream
    List<Operator<? extends OperatorDesc>> childOperators = parentRS.getChildOperators();
    int childIndex = childOperators.indexOf(mapJoinOp);
    childOperators.remove(childIndex);
    // at task startup
    if (mapJoinWork != null) {
        for (BaseWork myWork : mapJoinWork) {
            LOG.debug("adding dummy op to work " + myWork.getName() + " from MJ work: " + dummyOp);
            myWork.addDummyOp(dummyOp);
        }
    }
    if (context.linkChildOpWithDummyOp.containsKey(mapJoinOp)) {
        for (Operator<?> op : context.linkChildOpWithDummyOp.get(mapJoinOp)) {
            dummyOperators.add(op);
        }
    }
    context.linkChildOpWithDummyOp.put(mapJoinOp, dummyOperators);
    return true;
}