mariadb-columnstore-engine/dbcon/joblist/windowfunctionstep.cpp

/* Copyright (C) 2014 InfiniDB, Inc.
   Copyright (c) 2016-2020 MariaDB

   This program is free software; you can redistribute it and/or
   modify it under the terms of the GNU General Public License
   as published by the Free Software Foundation; version 2 of
   the License.

   This program is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   GNU General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with this program; if not, write to the Free Software
   Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
   MA 02110-1301, USA. */

//  $Id: windowfunctionstep.cpp 9681 2013-07-11 22:58:05Z xlou $

// #define NDEBUG
#include <cassert>
#include <sstream>
#include <iomanip>
using namespace std;

#include <boost/algorithm/string.hpp>  //  to_upper_copy
#include <boost/shared_ptr.hpp>

#include <boost/thread.hpp>
#include <boost/uuid/uuid_io.hpp>
using namespace boost;

#include "atomicops.h"
using namespace atomicops;

#include "loggingid.h"
#include "errorcodes.h"
#include "idberrorinfo.h"
using namespace logging;

#include "configcpp.h"
using namespace config;

#include "calpontselectexecutionplan.h"
#include "calpontsystemcatalog.h"
#include "aggregatecolumn.h"
#include "arithmeticcolumn.h"
#include "constantcolumn.h"
#include "functioncolumn.h"
#include "pseudocolumn.h"
#include "simplefilter.h"
#include "windowfunctioncolumn.h"
using namespace execplan;

#include "../../utils/windowfunction/windowfunction.h"
#include "../../utils/windowfunction/windowfunctiontype.h"
#include "../../utils/windowfunction/framebound.h"
#include "../../utils/windowfunction/frameboundrange.h"
#include "../../utils/windowfunction/frameboundrow.h"
#include "../../utils/windowfunction/windowframe.h"
using namespace windowfunction;

#include "rowgroup.h"
using namespace rowgroup;

using namespace ordering;

#include "funcexp.h"
using namespace funcexp;

#include "querytele.h"
using namespace querytele;

#include "jlf_common.h"
#include "jobstep.h"
#include "windowfunctionstep.h"
using namespace joblist;

#include "checks.h"

namespace
{
uint64_t getColumnIndex(const SRCP& c, const map<uint64_t, uint64_t>& m, JobInfo& jobInfo)
{
  uint64_t key = getTupleKey(jobInfo, c, true);
  const SimpleColumn* sc = dynamic_cast<const SimpleColumn*>(c.get());

  if (sc != NULL && !sc->schemaName().empty())
  {
    // special handling for dictionary
    CalpontSystemCatalog::ColType ct = sc->colType();

    // XXX use this before connector sets colType in sc correctly.
    //    type of pseudo column is set by connector
    if (!(dynamic_cast<const PseudoColumn*>(sc)))
    {
      ct = jobInfo.csc->colType(sc->oid());
      ct.charsetNumber = sc->colType().charsetNumber;
    }

    // X
    CalpontSystemCatalog::OID dictOid = isDictCol(ct);
    string alias(extractTableAlias(sc));

    if (dictOid > 0)
    {
      TupleInfo ti = setTupleInfo(ct, dictOid, jobInfo, tableOid(sc, jobInfo.csc), sc, alias);
      key = ti.key;
    }
  }

  map<uint64_t, uint64_t>::const_iterator j = m.find(key);

  if (j == m.end())
  {
    string name = jobInfo.keyInfo->tupleKeyToName[key];
    cerr << name << " is not in tuple, key=" << key << endl;
    throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_COLUMN_MISSING, name), ERR_WF_COLUMN_MISSING);
  }

  return j->second;
}

}  // namespace

namespace joblist
{
WindowFunctionStep::WindowFunctionStep(const JobInfo& jobInfo)
 : JobStep(jobInfo)
 , fRunner(0)
 , fCatalog(jobInfo.csc)
 , fRowsReturned(0)
 , fEndOfResult(false)
 , fIsSelect(true)
 , fUseSSMutex(false)
 , fUseUFMutex(false)
 , fInputDL(NULL)
 , fOutputDL(NULL)
 , fInputIterator(-1)
 , fOutputIterator(-1)
 , fFunctionCount(0)
 , fTotalThreads(1)
 , fNextIndex(0)
 , fMemUsage(0)
 , fRm(jobInfo.rm)
 , fSessionMemLimit(jobInfo.umMemLimit)
{
  fTotalThreads = fRm->windowFunctionThreads();
  fExtendedInfo = "WFS: ";
  fQtc.stepParms().stepType = StepTeleStats::T_WFS;
}

WindowFunctionStep::~WindowFunctionStep()
{
  if (fMemUsage > 0)
    fRm->returnMemory(fMemUsage, fSessionMemLimit);
}

void WindowFunctionStep::run()
{
  if (fInputJobStepAssociation.outSize() == 0)
    throw logic_error("No input data list for window function step.");

  fInputDL = fInputJobStepAssociation.outAt(0)->rowGroupDL();

  if (fInputDL == NULL)
    throw logic_error("Input is not a RowGroup data list in window function step.");

  fInputIterator = fInputDL->getIterator();

  if (fOutputJobStepAssociation.outSize() == 0)
    throw logic_error("No output data list for window function step.");

  fOutputDL = fOutputJobStepAssociation.outAt(0)->rowGroupDL();

  if (fOutputDL == NULL)
    throw logic_error("Output of window function step is not a RowGroup data list.");

  if (fDelivery == true)
  {
    fOutputIterator = fOutputDL->getIterator();
  }

  fRunner = jobstepThreadPool.invoke(Runner(this));
}

void WindowFunctionStep::join()
{
  if (fRunner)
    jobstepThreadPool.join(fRunner);
}

uint32_t WindowFunctionStep::nextBand(messageqcpp::ByteStream& bs)
{
  RGData rgDataOut;
  bool more = false;
  uint32_t rowCount = 0;

  try
  {
    bs.restart();

    more = fOutputDL->next(fOutputIterator, &rgDataOut);

    if (more && !cancelled())
    {
      fRowGroupDelivered.setData(&rgDataOut);
      fRowGroupDelivered.serializeRGData(bs);
      rowCount = fRowGroupDelivered.getRowCount();
    }
    else
    {
      while (more)
        more = fOutputDL->next(fOutputIterator, &rgDataOut);

      fEndOfResult = true;
    }
  }
  catch (...)
  {
    handleException(std::current_exception(), logging::ERR_IN_DELIVERY, logging::ERR_WF_DATA_SET_TOO_BIG,
                    "WindowFunctionStep::nextBand()");
    while (more)
      more = fOutputDL->next(fOutputIterator, &rgDataOut);
    fEndOfResult = true;
  }

  if (fEndOfResult)
  {
    // send an empty / error band
    rgDataOut.reinit(fRowGroupDelivered, 0);
    fRowGroupDelivered.setData(&rgDataOut);
    fRowGroupDelivered.resetRowGroup(0);
    fRowGroupDelivered.setStatus(status());
    fRowGroupDelivered.serializeRGData(bs);
  }

  return rowCount;
}

void WindowFunctionStep::setOutputRowGroup(const RowGroup& /*rg*/)
{
  idbassert(0);
}

const RowGroup& WindowFunctionStep::getOutputRowGroup() const
{
  return fRowGroupOut;
}

const RowGroup& WindowFunctionStep::getDeliveredRowGroup() const
{
  return fRowGroupDelivered;
}

void WindowFunctionStep::deliverStringTableRowGroup(bool b)
{
  fRowGroupOut.setUseStringTable(b);
  fRowGroupDelivered.setUseStringTable(b);
}

bool WindowFunctionStep::deliverStringTableRowGroup() const
{
  idbassert(fRowGroupOut.usesStringTable() == fRowGroupDelivered.usesStringTable());
  return fRowGroupDelivered.usesStringTable();
}

const string WindowFunctionStep::toString() const
{
  ostringstream oss;
  oss << "WindowFunctionStep   ses:" << fSessionId << " txn:" << fTxnId << " st:" << fStepId;

  oss << " in:";

  for (unsigned i = 0; i < fInputJobStepAssociation.outSize(); i++)
    oss << fInputJobStepAssociation.outAt(i);

  if (fOutputJobStepAssociation.outSize() > 0)
  {
    oss << " out:";

    for (unsigned i = 0; i < fOutputJobStepAssociation.outSize(); i++)
      oss << fOutputJobStepAssociation.outAt(i);
  }

  return oss.str();
}

void WindowFunctionStep::AddSimplColumn(const vector<SimpleColumn*>& scs, JobInfo& jobInfo)
{
  // append the simple columns if not already projected
  set<UniqId> scProjected;

  for (RetColsVector::iterator i = jobInfo.projectionCols.begin(); i != jobInfo.projectionCols.end(); i++)
  {
    SimpleColumn* sc = dynamic_cast<SimpleColumn*>(i->get());

    if (sc != NULL)
    {
      if (sc->schemaName().empty())
        sc->oid(joblist::tableOid(sc, jobInfo.csc) + 1 + sc->colPosition());

      scProjected.insert(UniqId(sc));
    }
  }

  for (vector<SimpleColumn*>::const_iterator i = scs.begin(); i != scs.end(); i++)
  {
    if (scProjected.find(UniqId(*i)) == scProjected.end())
    {
      jobInfo.windowDels.push_back(SRCP((*i)->clone()));
      // MCOL-3343 Enable this if we decide to allow Window Functions to run with
      // aggregates with no group by. MariaDB allows this. Nobody else in the world does.
      // There will be more work to get it to function if we try this.
      //            jobInfo.windowSet.insert(getTupleKey(jobInfo, *i, true));
      scProjected.insert(UniqId(*i));
    }
  }
}

void WindowFunctionStep::checkWindowFunction(CalpontSelectExecutionPlan* csep, JobInfo& jobInfo)
{
  // window functions in select clause, selected or in expression
  jobInfo.windowDels = jobInfo.deliveredCols;

  for (RetColsVector::iterator i = jobInfo.windowDels.begin(); i < jobInfo.windowDels.end(); i++)
  {
    const vector<WindowFunctionColumn*>& wcl = (*i)->windowfunctionColumnList();
    RetColsVector wcList;

    for (vector<WindowFunctionColumn*>::const_iterator j = wcl.begin(); j != wcl.end(); j++)
      wcList.push_back(SRCP((*j)->clone()));

    if (!wcList.empty())
    {
      jobInfo.windowExps.push_back(*i);
      jobInfo.windowSet.insert(getTupleKey(jobInfo, *i, true));
    }

    if (dynamic_cast<WindowFunctionColumn*>(i->get()) != NULL)
    {
      jobInfo.windowCols.push_back(*i);
      jobInfo.windowSet.insert(getTupleKey(jobInfo, *i, true));
    }
    else if (!wcList.empty())
    {
      jobInfo.windowCols.insert(jobInfo.windowCols.end(), wcList.begin(), wcList.end());

      for (RetColsVector::const_iterator k = wcList.begin(); k < wcList.end(); k++)
      {
        jobInfo.windowSet.insert(getTupleKey(jobInfo, *k, true));
      }
    }
  }

  // window functions in order by clause
  const CalpontSelectExecutionPlan::OrderByColumnList& orderByCols = csep->orderByCols();
  RetColsVector wcInOrderby;

  for (uint64_t i = 0; i < orderByCols.size(); i++)
  {
    if (orderByCols[i]->orderPos() == (uint64_t)(-1))
    {
      WindowFunctionColumn* wc = dynamic_cast<WindowFunctionColumn*>(orderByCols[i].get());
      const vector<WindowFunctionColumn*>& wcl = orderByCols[i]->windowfunctionColumnList();
      RetColsVector wcList;

      for (vector<WindowFunctionColumn*>::const_iterator j = wcl.begin(); j != wcl.end(); j++)
        wcList.push_back(SRCP((*j)->clone()));

      if (wc == NULL && wcList.empty())
        continue;

      // an window function or expression of window functions
      wcInOrderby.push_back(orderByCols[i]);

      if (!wcList.empty())
      {
        jobInfo.windowExps.push_back(orderByCols[i]);
        jobInfo.windowSet.insert(getTupleKey(jobInfo, orderByCols[i], true));
      }

      if (dynamic_cast<WindowFunctionColumn*>(orderByCols[i].get()) != NULL)
      {
        jobInfo.windowCols.push_back(orderByCols[i]);
        jobInfo.windowSet.insert(getTupleKey(jobInfo, orderByCols[i], true));
      }
      else if (!wcList.empty())
      {
        jobInfo.windowCols.insert(jobInfo.windowCols.end(), wcList.begin(), wcList.end());

        for (RetColsVector::const_iterator k = wcList.begin(); k < wcList.end(); k++)
        {
          jobInfo.windowSet.insert(getTupleKey(jobInfo, *k, true));
        }
      }
    }
  }

  // no window function involved in the query
  if (jobInfo.windowCols.empty())
    return;

  // Add in the non-window side of arithmetic columns and functions
  for (uint64_t i = 0; i < jobInfo.windowExps.size(); i++)
  {
    const ArithmeticColumn* ac = dynamic_cast<const ArithmeticColumn*>(jobInfo.windowExps[i].get());
    const FunctionColumn* fc = dynamic_cast<const FunctionColumn*>(jobInfo.windowExps[i].get());

    if (ac != NULL && ac->windowfunctionColumnList().size() > 0)
    {
      AddSimplColumn(ac->simpleColumnList(), jobInfo);
    }
    else if (fc != NULL && fc->windowfunctionColumnList().size() > 0)
    {
      AddSimplColumn(fc->simpleColumnList(), jobInfo);
    }
  }
  // reconstruct the delivered column list with auxiliary columns
  set<uint64_t> colSet;
  jobInfo.deliveredCols.resize(0);

  for (RetColsVector::iterator i = jobInfo.windowDels.begin(); i < jobInfo.windowDels.end(); i++)
  {
    jobInfo.deliveredCols.push_back(*i);
    uint64_t key = getTupleKey(jobInfo, *i, true);

    // TODO: remove duplicates in select clause
    colSet.insert(key);
  }

  // add window columns in orderby
  for (RetColsVector::iterator i = wcInOrderby.begin(); i < wcInOrderby.end(); i++)
  {
    jobInfo.deliveredCols.push_back(*i);
    uint64_t key = getTupleKey(jobInfo, *i, true);
    colSet.insert(key);
  }

  // MCOL-3435 We haven't yet checked for aggregate, but we need to know
  bool hasAggregation = false;
  for (uint64_t i = 0; i < jobInfo.deliveredCols.size(); i++)
  {
    if (dynamic_cast<AggregateColumn*>(jobInfo.deliveredCols[i].get()) != NULL)
    {
      hasAggregation = true;
      break;
    }
  }

  // add non-duplicate auxiliary columns
  RetColsVector colsInAf;

  for (RetColsVector::iterator i = jobInfo.windowCols.begin(); i < jobInfo.windowCols.end(); i++)
  {
    uint64_t key = getTupleKey(jobInfo, *i, true);

    if (colSet.find(key) == colSet.end())
      jobInfo.deliveredCols.push_back(*i);

    RetColsVector columns = dynamic_cast<WindowFunctionColumn*>(i->get())->getColumnList();

    for (RetColsVector::iterator j = columns.begin(); j < columns.end(); j++)
    {
      if (dynamic_cast<ConstantColumn*>(j->get()) != NULL)
        continue;

      key = getTupleKey(jobInfo, *j, true);

      if (colSet.find(key) == colSet.end())
      {
        jobInfo.deliveredCols.push_back(*j);
        // MCOL-3435 Allow Window Functions to run with aggregates with
        // no group by by inserting a group by for window parameters.
        if (hasAggregation)
        {
          // If an argument is an AggregateColumn, don't group by it.
          if (dynamic_cast<AggregateColumn*>(j->get()) == NULL)
          {
            bool bFound = false;
            for (std::vector<SRCP>::iterator igpc = csep->groupByCols().begin();
                 igpc < csep->groupByCols().end(); ++igpc)
            {
              if ((*igpc)->alias() == (*j)->alias())
              {
                bFound = true;
                break;
              }
            }
            if (!bFound)
            {
              csep->groupByCols().push_back(*j);
            }
          }
        }
      }

      colSet.insert(key);
    }
  }

  // for handling order by and limit in outer query
  jobInfo.wfqLimitStart = csep->limitStart();
  jobInfo.wfqLimitCount = csep->limitNum();
  csep->limitStart(0);
  csep->limitNum(-1);

  if (csep->orderByCols().size() > 0)
  {
    jobInfo.wfqOrderby = csep->orderByCols();
    csep->orderByCols().clear();

    // add order by columns
    for (RetColsVector::iterator i = jobInfo.wfqOrderby.begin(); i < jobInfo.wfqOrderby.end(); i++)
    {
      if (dynamic_cast<ConstantColumn*>(i->get()) != NULL)
        continue;

      uint64_t key = getTupleKey(jobInfo, *i, true);

      if (colSet.find(key) == colSet.end())
        jobInfo.deliveredCols.push_back(*i);

      colSet.insert(key);
    }
  }
}

SJSTEP WindowFunctionStep::makeWindowFunctionStep(SJSTEP& step, JobInfo& jobInfo)
{
  // create a window function step
  WindowFunctionStep* ws = new WindowFunctionStep(jobInfo);

  // connect to the feeding step
  JobStepAssociation jsa;
  AnyDataListSPtr spdl(new AnyDataList());
  RowGroupDL* dl = new RowGroupDL(1, jobInfo.fifoSize);
  dl->OID(execplan::CNX_VTABLE_ID);
  spdl->rowGroupDL(dl);
  jsa.outAdd(spdl);
  ws->inputAssociation(jsa);
  ws->stepId(step->stepId() + 1);
  step->outputAssociation(jsa);

  AnyDataListSPtr spdlOut(new AnyDataList());
  RowGroupDL* dlOut = new RowGroupDL(1, jobInfo.fifoSize);
  dlOut->OID(CNX_VTABLE_ID);
  spdlOut->rowGroupDL(dlOut);
  JobStepAssociation jsaOut;
  jsaOut.outAdd(spdlOut);
  ws->outputAssociation(jsaOut);

  // configure the rowgroups and index mapping
  TupleDeliveryStep* ds = dynamic_cast<TupleDeliveryStep*>(step.get());
  idbassert(ds != NULL);
  ws->initialize(ds->getDeliveredRowGroup(), jobInfo);

  // restore the original delivery coloumns
  jobInfo.deliveredCols = jobInfo.windowDels;
  jobInfo.nonConstDelCols.clear();

  for (RetColsVector::iterator i = jobInfo.windowDels.begin(); i < jobInfo.windowDels.end(); i++)
  {
    if (NULL == dynamic_cast<const ConstantColumn*>(i->get()))
      jobInfo.nonConstDelCols.push_back(*i);
  }

  return SJSTEP(ws);
}

void WindowFunctionStep::initialize(const RowGroup& rg, JobInfo& jobInfo)
{
  if (jobInfo.trace)
    cout << "Input to WindowFunctionStep: " << rg.toString() << endl;

  // query type decides the output by dbroot or partition
  // @bug 5631. Insert select should be treated as select
  fIsSelect = (jobInfo.queryType == "SELECT" || jobInfo.queryType == "INSERT_SELECT");

  // input row meta data
  fRowGroupIn = rg;
  fRowGroupIn.initRow(&fRowIn);

  // make an input map(id, index)
  map<uint64_t, uint64_t> colIndexMap;
  uint64_t colCntIn = rg.getColumnCount();
  const vector<uint32_t>& pos = rg.getOffsets();
  const vector<uint32_t>& oids = rg.getOIDs();
  const vector<uint32_t>& keys = rg.getKeys();
  const vector<CalpontSystemCatalog::ColDataType>& types = rg.getColTypes();
  const vector<uint32_t>& csNums = rg.getCharsetNumbers();
  const vector<uint32_t>& scales = rg.getScale();
  const vector<uint32_t>& precisions = rg.getPrecision();

  for (uint64_t i = 0; i < colCntIn; i++)
    colIndexMap.insert(make_pair(keys[i], i));

  // @bug6065, window functions that will update string table
  int64_t wfsUpdateStringTable = 0;
  int64_t wfsUserFunctionCount = 0;

  for (RetColsVector::iterator i = jobInfo.windowCols.begin(); i < jobInfo.windowCols.end(); i++)
  {
    bool isUDAF = false;
    // window function type
    WindowFunctionColumn* wc = dynamic_cast<WindowFunctionColumn*>(i->get());
    uint64_t ridx = getColumnIndex(*i, colIndexMap, jobInfo);  // result index
    // @bug6065, window functions that will update string table
    {
      CalpontSystemCatalog::ColType rt = wc->resultType();

      if ((types[ridx] == CalpontSystemCatalog::CHAR || types[ridx] == CalpontSystemCatalog::VARCHAR ||
           types[ridx] == CalpontSystemCatalog::TEXT || types[ridx] == CalpontSystemCatalog::VARBINARY ||
           types[ridx] == CalpontSystemCatalog::BLOB) &&
          rg.getColumnWidth(ridx) >= jobInfo.stringTableThreshold)
      {
        ++wfsUpdateStringTable;
      }
    }

    //		if (boost::iequals(wc->functionName(),"UDAF_FUNC")
    if (wc->functionName() == "UDAF_FUNC")
    {
      isUDAF = true;
      ++wfsUserFunctionCount;
    }

    vector<int64_t> fields;
    fields.push_back(ridx);  // result
    const RetColsVector& parms = wc->functionParms();

    for (uint64_t i = 0; i < parms.size(); i++)  // arguments
    {
      // skip constant column
      if (dynamic_cast<const ConstantColumn*>(parms[i].get()) == NULL)
        fields.push_back(getColumnIndex(parms[i], colIndexMap, jobInfo));
      else
        fields.push_back(-1);
    }

    // partition & order by
    const RetColsVector& partitions = wc->partitions();
    vector<uint64_t> eqIdx;
    vector<uint64_t> peerIdx;
    vector<IdbSortSpec> sorts;

    for (uint64_t i = 0; i < partitions.size(); i++)
    {
      // skip constant column
      if (dynamic_cast<const ConstantColumn*>(partitions[i].get()) != NULL)
        continue;

      // get column index
      uint64_t idx = getColumnIndex(partitions[i], colIndexMap, jobInfo);
      eqIdx.push_back(idx);
      sorts.push_back(IdbSortSpec(idx, partitions[i]->asc(), partitions[i]->nullsFirst()));
    }

    const RetColsVector& orders = wc->orderBy().fOrders;

    for (uint64_t i = 0; i < orders.size(); i++)
    {
      // skip constant column
      if (dynamic_cast<const ConstantColumn*>(orders[i].get()) != NULL)
        continue;

      // get column index
      uint64_t idx = getColumnIndex(orders[i], colIndexMap, jobInfo);
      peerIdx.push_back(idx);
      sorts.push_back(IdbSortSpec(idx, orders[i]->asc(), orders[i]->nullsFirst()));
    }

    // functors for sorting
    boost::shared_ptr<EqualCompData> parts(new EqualCompData(eqIdx, rg));
    boost::shared_ptr<OrderByData> orderbys(new OrderByData(sorts, rg));
    boost::shared_ptr<EqualCompData> peers(new EqualCompData(peerIdx, rg));

    // column type for functor templates
    int ct = 0;

    if (isUDAF)
    {
      ct = wc->getUDAFContext().getResultType();
    }
    // make sure index is in range
    else if (fields.size() > 1 && fields[1] >= 0 && static_cast<uint64_t>(fields[1]) < types.size())
      ct = types[fields[1]];

    // workaround for functions using "within group (order by)" syntax
    string fn = boost::to_upper_copy(wc->functionName());

    if ((fn == "MEDIAN" || fn == "PERCENTILE_CONT" || fn == "PERCENTILE_DISC") &&
        utils::is_nonnegative(peerIdx[0]) && peerIdx[0] < types.size())
      ct = types[peerIdx[0]];

    // create the functor based on function name
    boost::shared_ptr<WindowFunctionType> func = WindowFunctionType::makeWindowFunction(fn, ct, wc);

    // parse parms after peer and fields are set
    // functions may need to set order column index
    func->peer(peers);
    func->fieldIndex(fields);
    func->parseParms(parms);

    // window frame
    const WF_Frame& frame = wc->orderBy().fFrame;
    int frameUnit = (frame.fIsRange) ? WF__FRAME_RANGE : WF__FRAME_ROWS;

    if (frame.fStart.fFrame == WF_UNBOUNDED_PRECEDING && frame.fEnd.fFrame == WF_UNBOUNDED_FOLLOWING)
      frameUnit = WF__FRAME_ROWS;

    boost::shared_ptr<FrameBound> upper =
        parseFrameBound(frame.fStart, colIndexMap, orders, peers, jobInfo, !frame.fIsRange, true);
    boost::shared_ptr<FrameBound> lower =
        parseFrameBound(frame.fEnd, colIndexMap, orders, peers, jobInfo, !frame.fIsRange, false);
    boost::shared_ptr<WindowFrame> windows(new WindowFrame(frameUnit, upper, lower));
    func->frameUnit(frameUnit);

    // add to the function list
    fFunctions.push_back(
        boost::shared_ptr<WindowFunction>(new WindowFunction(func, parts, orderbys, windows, rg, fRowIn)));
    fFunctionCount++;
  }

  // initialize window function expresssions
  fExpression = jobInfo.windowExps;

  for (RetColsVector::iterator i = fExpression.begin(); i < fExpression.end(); i++)
  {
    // output index
    (*i)->outputIndex(getColumnIndex(*i, colIndexMap, jobInfo));

    // map the input indices
    const vector<SimpleColumn*>& scols = (*i)->simpleColumnList();

    for (vector<SimpleColumn*>::const_iterator j = scols.begin(); j != scols.end(); j++)
    {
      uint64_t key = getTupleKey(jobInfo, *j);
      CalpontSystemCatalog::OID dictOid = joblist::isDictCol((*j)->colType());

      if (dictOid > 0)
      {
        key = jobInfo.keyInfo->dictKeyMap[key];
      }

      map<uint64_t, uint64_t>::iterator k = colIndexMap.find(key);

      if (k != colIndexMap.end())
      {
        (*j)->inputIndex(k->second);
      }
      else
      {
        string name = jobInfo.keyInfo->tupleKeyToName[key];
        cerr << name << " is not in tuple, key=" << key << endl;
        throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_COLUMN_MISSING, name),
                        ERR_WF_COLUMN_MISSING);
      }
    }

    ArithmeticColumn* ac = dynamic_cast<ArithmeticColumn*>((*i).get());
    FunctionColumn* fc = dynamic_cast<FunctionColumn*>((*i).get());

    if (ac != NULL)
    {
      updateWindowCols(ac->expression(), colIndexMap, jobInfo);
    }
    else if (fc != NULL)
    {
      //			RetColsVector wcList = fc->windowfunctionColumnList();
      //			for (RetColsVector::iterator j = wcList.begin(); j != wcList.end(); j++)
      //				(*j)->inputIndex(getColumnIndex(*j, colIndexMap, jobInfo));
      funcexp::FunctionParm parms = fc->functionParms();

      for (vector<execplan::SPTP>::iterator j = parms.begin(); j < parms.end(); j++)
        updateWindowCols(j->get(), colIndexMap, jobInfo);
    }
  }

  // order by part
  if (jobInfo.wfqOrderby.size() > 0)
  {
    // query order by
    vector<uint64_t> eqIdx;
    vector<IdbSortSpec> sorts;
    const RetColsVector& orderby = jobInfo.wfqOrderby;

    for (uint64_t i = 0; i < orderby.size(); i++)
    {
      // skip constant column
      if (dynamic_cast<const ConstantColumn*>(orderby[i].get()) != NULL)
        continue;

      // get column index
      uint64_t idx = getColumnIndex(orderby[i], colIndexMap, jobInfo);
      sorts.push_back(IdbSortSpec(idx, orderby[i]->asc(), orderby[i]->nullsFirst()));
    }

    fQueryOrderBy.reset(new OrderByData(sorts, rg));
  }

  // limit part
  fQueryLimitStart = jobInfo.wfqLimitStart;
  fQueryLimitCount = jobInfo.wfqLimitCount;

  // fix the delivered rowgroup data
  vector<uint64_t> delColIdx;

  for (RetColsVector::iterator i = jobInfo.windowDels.begin(); i < jobInfo.windowDels.end(); i++)
  {
    // find the none constantant columns in the deliver
    // leave constants to annexstep for now.
    if (dynamic_cast<const ConstantColumn*>((*i).get()) != NULL)
      continue;

    delColIdx.push_back(getColumnIndex(*i, colIndexMap, jobInfo));
  }

  size_t retColCount = delColIdx.size();
  vector<uint32_t> pos1;
  vector<uint32_t> oids1;
  vector<uint32_t> keys1;
  vector<uint32_t> scales1;
  vector<uint32_t> precisions1;
  vector<CalpontSystemCatalog::ColDataType> types1;
  vector<uint32_t> csNums1;
  pos1.push_back(2);

  for (size_t i = 0; i < retColCount; i++)
  {
    size_t j = delColIdx[i];
    pos1.push_back(pos1[i] + (pos[j + 1] - pos[j]));
    oids1.push_back(oids[j]);
    keys1.push_back(keys[j]);
    scales1.push_back(scales[j]);
    precisions1.push_back(precisions[j]);
    types1.push_back(types[j]);
    csNums1.push_back(csNums[j]);
  }

  fRowGroupDelivered = RowGroup(retColCount, pos1, oids1, keys1, types1, csNums1, scales1, precisions1,
                                jobInfo.stringTableThreshold);

  if (jobInfo.trace)
    cout << "delivered RG: " << fRowGroupDelivered.toString() << endl << endl;

  if (wfsUpdateStringTable > 1)
    fUseSSMutex = true;

  if (wfsUserFunctionCount > 1)
    fUseUFMutex = true;

  fRowGroupOut = fRowGroupDelivered;
}

void WindowFunctionStep::execute()
{
  RGData rgData;
  Row row;
  fRowGroupIn.initRow(&row);
  bool more = fInputDL->next(fInputIterator, &rgData);
  uint64_t i = 0;  // for RowGroup index in the fInRowGroupData

  if (traceOn())
    dlTimes.setFirstReadTime();

  {
    StepTeleStats sts(fQueryUuid, fStepUuid, StepTeleStats::ST_START, 1);
    postStepStartTele(sts);
  }

  try
  {
    while (more && !cancelled())
    {
      fRowGroupIn.setData(&rgData);
      fRowGroupIn.getRow(0, &row);
      uint64_t rowCnt = fRowGroupIn.getRowCount();

      if (rowCnt > 0)
      {
        fInRowGroupData.push_back(rgData);
        uint64_t memAdd = fRowGroupIn.getSizeWithStrings() + rowCnt * sizeof(RowPosition);

        if (fRm->getMemory(memAdd, fSessionMemLimit) == false)
          throw IDBExcept(ERR_WF_DATA_SET_TOO_BIG);

        fMemUsage += memAdd;

        for (uint64_t j = 0; j < rowCnt; ++j)
        {
          if (i > 0x0000FFFFFFFFFFFFULL || j > 0x000000000000FFFFULL)
            throw IDBExcept(ERR_WF_DATA_SET_TOO_BIG);

          fRows.push_back(RowPosition(i, j));
          row.nextRow();
        }

        //@bug6065, make StringStore::storeString() thread safe, default to false.
        rgData.useStoreStringMutex(fUseSSMutex);
        // For the User Data of UDAnF
        rgData.useUserDataMutex(fUseUFMutex);

        // window function does not change row count
        fRowsReturned += rowCnt;

        i++;
      }

      more = fInputDL->next(fInputIterator, &rgData);
    }
  }  // try
  catch (...)
  {
    handleException(std::current_exception(), logging::ERR_READ_INPUT_DATALIST,
                    logging::ERR_WF_DATA_SET_TOO_BIG, "WindowFunctionStep::execute()");
  }

  if (traceOn())
    dlTimes.setLastReadTime();

  // no need for the window function if aborted or result set is empty.
  if (cancelled() || fRows.size() == 0)
  {
    while (more)
      more = fInputDL->next(fInputIterator, &rgData);

    fOutputDL->endOfInput();

    StepTeleStats sts(fQueryUuid, fStepUuid, StepTeleStats::ST_SUMMARY, 1, 1, fRowsReturned);
    postStepSummaryTele(sts);

    if (traceOn())
    {
      dlTimes.setEndOfInputTime();
      printCalTrace();
    }

    return;
  }

  // got something to work on
  try
  {
    if (fFunctionCount == 1)
    {
      doFunction();
    }
    else
    {
      if (fTotalThreads > fFunctionCount)
        fTotalThreads = fFunctionCount;

      fFunctionThreads.clear();
      fFunctionThreads.reserve(fTotalThreads);

      for (uint64_t i = 0; i < fTotalThreads && !cancelled(); i++)
        fFunctionThreads.push_back(jobstepThreadPool.invoke(WFunction(this)));

      // If cancelled, not all threads are started.
      jobstepThreadPool.join(fFunctionThreads);
    }

    if (!(cancelled()))
    {
      if (fIsSelect)
        doPostProcessForSelect();
      else
        doPostProcessForDml();
    }
  }
  catch (...)
  {
    handleException(std::current_exception(), logging::ERR_EXECUTE_WINDOW_FUNCTION,
                    logging::ERR_WF_DATA_SET_TOO_BIG, "WindowFunctionStep::execute()");
  }

  fOutputDL->endOfInput();

  StepTeleStats sts(fQueryUuid, fStepUuid, StepTeleStats::ST_SUMMARY, 1, 1, fRowsReturned);
  postStepSummaryTele(sts);

  if (traceOn())
  {
    dlTimes.setEndOfInputTime();
    printCalTrace();
  }

  return;
}

uint64_t WindowFunctionStep::nextFunctionIndex()
{
  uint64_t idx = atomicInc(&fNextIndex);

  // return index in the function array
  return --idx;
}

void WindowFunctionStep::doFunction()
{
  uint64_t i = 0;

  try
  {
    while (((i = nextFunctionIndex()) < fFunctionCount) && !cancelled())
    {
      uint64_t memAdd = fRows.size() * sizeof(RowPosition);

      if (fRm->getMemory(memAdd, fSessionMemLimit) == false)
        throw IDBExcept(ERR_WF_DATA_SET_TOO_BIG);

      fMemUsage += memAdd;

      fFunctions[i]->setCallback(this, i);
      (*fFunctions[i].get())();
    }
  }
  catch (...)
  {
    handleException(std::current_exception(), logging::ERR_EXECUTE_WINDOW_FUNCTION,
                    logging::ERR_WF_DATA_SET_TOO_BIG, "WindowFunctionStep::doFunction()");
  }
}

void WindowFunctionStep::doPostProcessForSelect()
{
  FuncExp* fe = funcexp::FuncExp::instance();
  std::shared_ptr<int[]> mapping = makeMapping(fRowGroupIn, fRowGroupOut);
  Row rowIn, rowOut;
  fRowGroupIn.initRow(&rowIn);
  fRowGroupOut.initRow(&rowOut);
  RGData rgData;
  vector<RowPosition>& rowData = *(fFunctions.back()->fRowData.get());
  int64_t rowsLeft = rowData.size();
  int64_t rowsInRg = 0;
  int64_t rgCapacity = 0;

  int64_t begin = fQueryLimitStart;
  int64_t count = (fQueryLimitCount == (uint64_t)-1) ? rowsLeft : fQueryLimitCount;
  int64_t end = begin + count;
  end = (end < rowsLeft) ? end : rowsLeft;
  rowsLeft = (end > begin) ? (end - begin) : 0;

  if (fQueryOrderBy.get() != NULL)
    sort(rowData.begin(), rowData.size());

  for (int64_t i = begin; i < end; i++)
  {
    if (!rgData.hasRowData())
    {
      rgCapacity = ((rowsLeft > 8192) ? 8192 : rowsLeft);
      rowsLeft -= rgCapacity;
      rgData.reinit(fRowGroupOut, rgCapacity);

      fRowGroupOut.setData(&rgData);
      fRowGroupOut.resetRowGroup(0);
      fRowGroupOut.setDBRoot(0);  // not valid dbroot
      fRowGroupOut.getRow(0, &rowOut);
      rowsInRg = 0;
    }

    rowIn.setData(getPointer(rowData[i], fRowGroupIn, rowIn));

    // evaluate the window function expressions before apply mapping
    if (fExpression.size() > 0)
      fe->evaluate(rowIn, fExpression);

    applyMapping(mapping, rowIn, &rowOut);
    rowOut.nextRow();
    rowsInRg++;

    if (rowsInRg == rgCapacity)
    {
      fRowGroupOut.setRowCount(rowsInRg);
      fOutputDL->insert(rgData);
      rgData.clear();
    }
  }
}

void WindowFunctionStep::doPostProcessForDml()
{
  FuncExp* fe = funcexp::FuncExp::instance();
  std::shared_ptr<int[]> mapping = makeMapping(fRowGroupIn, fRowGroupOut);
  Row rowIn, rowOut;
  fRowGroupIn.initRow(&rowIn);
  fRowGroupOut.initRow(&rowOut);

  for (vector<RGData>::iterator i = fInRowGroupData.begin(); i < fInRowGroupData.end(); i++)
  {
    fRowGroupIn.setData(&(*i));
    RGData rgData = RGData(fRowGroupIn, fRowGroupIn.getRowCount());
    fRowGroupOut.setData(&rgData);
    // @bug 5631. reset rowgroup before the data is populated.
    fRowGroupOut.resetRowGroup(fRowGroupIn.getBaseRid());
    fRowGroupOut.setDBRoot(fRowGroupIn.getDBRoot());
    fRowGroupOut.setRowCount(fRowGroupIn.getRowCount());

    fRowGroupIn.getRow(0, &rowIn);
    fRowGroupOut.getRow(0, &rowOut);

    for (uint64_t i = 0; i < fRowGroupIn.getRowCount(); ++i)
    {
      // evaluate the window function expressions before apply mapping
      if (fExpression.size() > 0)
        fe->evaluate(rowIn, fExpression);

      applyMapping(mapping, rowIn, &rowOut);
      rowIn.nextRow();
      rowOut.nextRow();
    }

    // fRowGroupOut.resetRowGroup(fRowGroupIn.getBaseRid());
    // fRowGroupOut.setRowCount(fRowGroupIn.getRowCount());

    fOutputDL->insert(rgData);
  }
}

boost::shared_ptr<FrameBound> WindowFunctionStep::parseFrameBoundRows(const execplan::WF_Boundary& b,
                                                                      const map<uint64_t, uint64_t>& m,
                                                                      JobInfo& jobInfo)
{
  boost::shared_ptr<FrameBound> fb;

  if (b.fFrame == WF_CURRENT_ROW)
  {
    fb.reset(new FrameBoundRow(WF__CURRENT_ROW));
    return fb;
  }

  ConstantColumn* cc = dynamic_cast<ConstantColumn*>(b.fVal.get());

  if (cc != NULL)
  {
    Row dummy;
    bool isNull = false;
    int val = cc->getIntVal(dummy, isNull);

    if (val >= 0 && isNull == false)
    {
      int type = (b.fFrame == WF_PRECEDING) ? WF__CONSTANT_PRECEDING : WF__CONSTANT_FOLLOWING;
      fb.reset(new FrameBoundConstantRow(type, val));
    }
    else
    {
      string str("NULL");

      if (!isNull)
      {
        ostringstream oss;
        oss << val;
        str = oss.str();
      }

      throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_BOUND_OUT_OF_RANGE, str),
                      ERR_WF_BOUND_OUT_OF_RANGE);
    }
  }
  else
  {
    int type = (b.fFrame == WF_PRECEDING) ? WF__EXPRESSION_PRECEDING : WF__EXPRESSION_FOLLOWING;
    uint64_t id = getTupleKey(jobInfo, b.fVal);
    uint64_t idx = getColumnIndex(b.fVal, m, jobInfo);
    TupleInfo ti = getTupleInfo(id, jobInfo);

    switch (ti.dtype)
    {
      case execplan::CalpontSystemCatalog::TINYINT:
      case execplan::CalpontSystemCatalog::SMALLINT:
      case execplan::CalpontSystemCatalog::MEDINT:
      case execplan::CalpontSystemCatalog::INT:
      case execplan::CalpontSystemCatalog::BIGINT:
      case execplan::CalpontSystemCatalog::DECIMAL:
      {
        fb.reset(new FrameBoundExpressionRow<int64_t>(type, id, idx));
        break;
      }

      case execplan::CalpontSystemCatalog::DOUBLE:
      case execplan::CalpontSystemCatalog::UDOUBLE:
      {
        fb.reset(new FrameBoundExpressionRow<double>(type, id, idx));
        break;
      }

      case execplan::CalpontSystemCatalog::FLOAT:
      case execplan::CalpontSystemCatalog::UFLOAT:
      {
        fb.reset(new FrameBoundExpressionRow<float>(type, id, idx));
        break;
      }

      case execplan::CalpontSystemCatalog::UTINYINT:
      case execplan::CalpontSystemCatalog::USMALLINT:
      case execplan::CalpontSystemCatalog::UMEDINT:
      case execplan::CalpontSystemCatalog::UINT:
      case execplan::CalpontSystemCatalog::UBIGINT:
      case execplan::CalpontSystemCatalog::UDECIMAL:
      case execplan::CalpontSystemCatalog::DATE:
      case execplan::CalpontSystemCatalog::DATETIME:
      case execplan::CalpontSystemCatalog::TIME:
      case execplan::CalpontSystemCatalog::TIMESTAMP:
      {
        fb.reset(new FrameBoundExpressionRow<uint64_t>(type, id, idx));
        break;
      }

      default:
      {
        string str = windowfunction::colType2String[ti.dtype];
        throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_INVALID_BOUND_TYPE, str),
                        ERR_WF_INVALID_BOUND_TYPE);
        break;
      }
    }
  }

  return fb;
}

boost::shared_ptr<FrameBound> WindowFunctionStep::parseFrameBoundRange(const execplan::WF_Boundary& b,
                                                                       const map<uint64_t, uint64_t>& m,
                                                                       const vector<SRCP>& o,
                                                                       JobInfo& jobInfo)
{
  boost::shared_ptr<FrameBound> fb;

  if (b.fFrame == WF_CURRENT_ROW)
  {
    fb.reset(new FrameBoundRange(WF__CURRENT_ROW));
    return fb;
  }

  bool isConstant = false;
  bool isNull = false;
  Row dummy;
  ConstantColumn* cc = dynamic_cast<ConstantColumn*>(b.fVal.get());

  if (cc != NULL)
  {
    isConstant = true;
    double val = cc->getDoubleVal(dummy, isNull);

    if (val < 0 || isNull)
    {
      string str("NULL");

      if (!isNull)
      {
        ostringstream oss;
        oss << val;
        str = oss.str();
      }

      throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_BOUND_OUT_OF_RANGE, str),
                      ERR_WF_BOUND_OUT_OF_RANGE);
    }
  }

  int type = 0;
  vector<uint64_t> ids;
  vector<int> index;
  ids.push_back(getTupleKey(jobInfo, o[0]));
  index.push_back(getColumnIndex(o[0], m, jobInfo));

  if (isConstant)
  {
    type = (b.fFrame == WF_PRECEDING) ? WF__CONSTANT_PRECEDING : WF__CONSTANT_FOLLOWING;
    ids.push_back(-1);    // dummy, n/a for constant
    index.push_back(-1);  // dummy, n/a for constant
  }
  else
  {
    type = (b.fFrame == WF_PRECEDING) ? WF__EXPRESSION_PRECEDING : WF__EXPRESSION_FOLLOWING;
    ids.push_back(getTupleKey(jobInfo, b.fVal));
    index.push_back(getColumnIndex(b.fVal, m, jobInfo));
  }

  ids.push_back(getTupleKey(jobInfo, b.fBound));
  index.push_back(getColumnIndex(b.fBound, m, jobInfo));

  FrameBoundRange* fbr = NULL;
  TupleInfo ti = getTupleInfo(ids[0], jobInfo);
  bool asc = o[0]->asc();
  bool nlf = o[0]->nullsFirst();

  switch (ti.dtype)
  {
    case execplan::CalpontSystemCatalog::TINYINT:
    case execplan::CalpontSystemCatalog::SMALLINT:
    case execplan::CalpontSystemCatalog::MEDINT:
    case execplan::CalpontSystemCatalog::INT:
    case execplan::CalpontSystemCatalog::BIGINT:
    case execplan::CalpontSystemCatalog::DECIMAL:
    {
      if (isConstant)
      {
        int64_t v = cc->getIntVal(dummy, isNull);
        fbr = new FrameBoundConstantRange<int64_t>(type, asc, nlf, &v);
        fbr->isZero((v == 0));
      }
      else
      {
        fbr = new FrameBoundExpressionRange<int64_t>(type, asc, nlf);
      }

      break;
    }

    case execplan::CalpontSystemCatalog::DOUBLE:
    case execplan::CalpontSystemCatalog::UDOUBLE:
    {
      if (isConstant)
      {
        double v = cc->getDoubleVal(dummy, isNull);
        fbr = new FrameBoundConstantRange<double>(type, asc, nlf, &v);
        fbr->isZero((v == 0.0));
      }
      else
      {
        fbr = new FrameBoundExpressionRange<double>(type, asc, nlf);
      }

      break;
    }

    case execplan::CalpontSystemCatalog::FLOAT:
    case execplan::CalpontSystemCatalog::UFLOAT:
    {
      if (isConstant)
      {
        float v = cc->getFloatVal(dummy, isNull);
        fbr = new FrameBoundConstantRange<float>(type, asc, nlf, &v);
        fbr->isZero((v == 0.0));
      }
      else
      {
        fbr = new FrameBoundExpressionRange<float>(type, asc, nlf);
      }

      break;
    }

    case execplan::CalpontSystemCatalog::UTINYINT:
    case execplan::CalpontSystemCatalog::USMALLINT:
    case execplan::CalpontSystemCatalog::UMEDINT:
    case execplan::CalpontSystemCatalog::UINT:
    case execplan::CalpontSystemCatalog::UBIGINT:
    case execplan::CalpontSystemCatalog::UDECIMAL:
    case execplan::CalpontSystemCatalog::DATE:
    case execplan::CalpontSystemCatalog::DATETIME:
    case execplan::CalpontSystemCatalog::TIME:
    case execplan::CalpontSystemCatalog::TIMESTAMP:
    {
      if (isConstant)
      {
        uint64_t v = cc->getUintVal(dummy, isNull);
        fbr = new FrameBoundConstantRange<uint64_t>(type, asc, nlf, &v);
        fbr->isZero((v == 0));
      }
      else
      {
        fbr = new FrameBoundExpressionRange<uint64_t>(type, asc, nlf);
      }

      break;
    }

    default:
    {
      string str = windowfunction::colType2String[ti.dtype];
      throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_INVALID_BOUND_TYPE, str),
                      ERR_WF_INVALID_BOUND_TYPE);
      break;
    }
  }

  fbr->setTupleId(ids);
  fbr->setIndex(index);
  fb.reset(fbr);

  return fb;
}

boost::shared_ptr<FrameBound> WindowFunctionStep::parseFrameBound(const execplan::WF_Boundary& b,
                                                                  const map<uint64_t, uint64_t>& m,
                                                                  const vector<SRCP>& o,
                                                                  const boost::shared_ptr<EqualCompData>& p,
                                                                  JobInfo& j, bool rows, bool s)
{
  boost::shared_ptr<FrameBound> fb;

  switch (b.fFrame)
  {
    case WF_UNBOUNDED_PRECEDING:
    {
      fb.reset(new FrameBound(WF__UNBOUNDED_PRECEDING));
      break;
    }

    case WF_UNBOUNDED_FOLLOWING:
    {
      fb.reset(new FrameBound(WF__UNBOUNDED_FOLLOWING));
      break;
    }

    case WF_CURRENT_ROW:
    case WF_PRECEDING:
    case WF_FOLLOWING:
    {
      if (rows)
      {
        fb = parseFrameBoundRows(b, m, j);
      }
      else
      {
        fb = parseFrameBoundRange(b, m, o, j);
      }

      break;
    }

    default:  //  unknown
    {
      throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_UNKNOWN_BOUND, b.fFrame),
                      ERR_WF_UNKNOWN_BOUND);
      break;
    }
  }

  fb->peer(p);
  fb->start(s);

  return fb;
}

void WindowFunctionStep::updateWindowCols(ReturnedColumn* rc, const map<uint64_t, uint64_t>& m,
                                          JobInfo& jobInfo)
{
  if (rc == NULL)
    return;

  ArithmeticColumn* ac = dynamic_cast<ArithmeticColumn*>(rc);
  FunctionColumn* fc = dynamic_cast<FunctionColumn*>(rc);
  SimpleFilter* sf = dynamic_cast<SimpleFilter*>(rc);
  WindowFunctionColumn* wc = dynamic_cast<WindowFunctionColumn*>(rc);

  if (wc)
  {
    uint64_t key = getExpTupleKey(jobInfo, wc->expressionId());
    map<uint64_t, uint64_t>::const_iterator j = m.find(key);

    if (j == m.end())
    {
      string name = jobInfo.keyInfo->tupleKeyToName[key];
      cerr << name << " is not in tuple, key=" << key << endl;
      throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_COLUMN_MISSING, name), ERR_WF_COLUMN_MISSING);
    }

    wc->inputIndex(j->second);
  }
  else if (ac)
  {
    updateWindowCols(ac->expression(), m, jobInfo);
  }
  else if (fc)
  {
    funcexp::FunctionParm parms = fc->functionParms();

    for (vector<execplan::SPTP>::iterator i = parms.begin(); i < parms.end(); i++)
      updateWindowCols(i->get(), m, jobInfo);
  }
  else if (sf)
  {
    updateWindowCols(sf->lhs(), m, jobInfo);
    updateWindowCols(sf->rhs(), m, jobInfo);
  }
}

void WindowFunctionStep::updateWindowCols(ParseTree* pt, const map<uint64_t, uint64_t>& m, JobInfo& jobInfo)
{
  if (pt == NULL)
    return;

  updateWindowCols(pt->left(), m, jobInfo);
  updateWindowCols(pt->right(), m, jobInfo);

  TreeNode* tn = pt->data();
  ArithmeticColumn* ac = dynamic_cast<ArithmeticColumn*>(tn);
  FunctionColumn* fc = dynamic_cast<FunctionColumn*>(tn);
  SimpleFilter* sf = dynamic_cast<SimpleFilter*>(tn);
  WindowFunctionColumn* wc = dynamic_cast<WindowFunctionColumn*>(tn);

  if (wc)
  {
    uint64_t key = getExpTupleKey(jobInfo, wc->expressionId());
    map<uint64_t, uint64_t>::const_iterator j = m.find(key);

    if (j == m.end())
    {
      string name = jobInfo.keyInfo->tupleKeyToName[key];
      cerr << name << " is not in tuple, key=" << key << endl;
      throw IDBExcept(IDBErrorInfo::instance()->errorMsg(ERR_WF_COLUMN_MISSING, name), ERR_WF_COLUMN_MISSING);
    }

    wc->inputIndex(j->second);
  }
  else if (ac)
  {
    updateWindowCols(ac->expression(), m, jobInfo);
  }
  else if (fc)
  {
    funcexp::FunctionParm parms = fc->functionParms();

    for (vector<execplan::SPTP>::iterator i = parms.begin(); i < parms.end(); i++)
      updateWindowCols(i->get(), m, jobInfo);
  }
  else if (sf)
  {
    updateWindowCols(sf->lhs(), m, jobInfo);
    updateWindowCols(sf->rhs(), m, jobInfo);
  }
}

void WindowFunctionStep::sort(std::vector<RowPosition>::iterator v, uint64_t n)
{
  // recursive function termination condition.
  if (n < 2 || cancelled())
    return;

  RowPosition p = *(v + n / 2);                   // pivot value
  vector<RowPosition>::iterator l = v;            // low   address
  vector<RowPosition>::iterator h = v + (n - 1);  // high  address

  while (l <= h && !cancelled())
  {
    // Can use while here, but need check boundary and cancel status.
    if (fQueryOrderBy->operator()(getPointer(*l), getPointer(p)))
    {
      l++;
    }
    else if (fQueryOrderBy->operator()(getPointer(p), getPointer(*h)))
    {
      h--;
    }
    else
    {
      RowPosition t = *l;  // temp value for swap
      *l++ = *h;
      *h-- = t;
    }
  }

  sort(v, std::distance(v, h) + 1);
  sort(l, std::distance(l, v) + n);
}

void WindowFunctionStep::printCalTrace()
{
  time_t t = time(0);
  char timeString[50];
  ctime_r(&t, timeString);
  timeString[strlen(timeString) - 1] = '\0';
  ostringstream logStr;
  logStr << "ses:" << fSessionId << " st: " << fStepId << " finished at " << timeString
         << "; total rows returned-" << fRowsReturned << endl
         << "\t1st read " << dlTimes.FirstReadTimeString() << "; EOI " << dlTimes.EndOfInputTimeString()
         << "; runtime-" << JSTimeStamp::tsdiffstr(dlTimes.EndOfInputTime(), dlTimes.FirstReadTime())
         << "s;\n\tUUID " << uuids::to_string(fStepUuid) << endl
         << "\tJob completion status " << status() << endl;
  logEnd(logStr.str().c_str());
  fExtendedInfo += logStr.str();
  formatMiniStats();
}

void WindowFunctionStep::formatMiniStats()
{
  ostringstream oss;
  oss << "WFS " << "UM " << "- " << "- " << "- " << "- " << "- " << "- "
      << JSTimeStamp::tsdiffstr(dlTimes.EndOfInputTime(), dlTimes.FirstReadTime()) << " " << fRowsReturned
      << " ";
  fMiniInfo += oss.str();
}

}  // namespace joblist