RArrowDS.cxx

Go to the documentation of this file.

// Author: Giulio Eulisse CERN  2/2018
 
/*************************************************************************
 * Copyright (C) 1995-2018, Rene Brun and Fons Rademakers.               *
 * All rights reserved.                                                  *
 *                                                                       *
 * For the licensing terms see $ROOTSYS/LICENSE.                         *
 * For the list of contributors see $ROOTSYS/README/CREDITS.             *
 *************************************************************************/
 
// clang-format off
/** \class ROOT::RDF::RArrowDS
    \ingroup dataframe
    \brief RDataFrame data source class to interface with Apache Arrow.
 
The RArrowDS implements a proxy RDataSource to be able to use Apache Arrow
tables with RDataFrame.
 
A RDataFrame that adapts an arrow::Table class can be constructed using the factory method
ROOT::RDF::MakeArrowDataFrame, which accepts one parameter:
1. An arrow::Table smart pointer.
 
The types of the columns are derived from the types in the associated
arrow::Schema.
 
*/
// clang-format on
 
#include <ROOT/RDF/Utils.hxx>
#include <ROOT/TSeq.hxx>
#include <ROOT/RArrowDS.hxx>
#include <ROOT/RMakeUnique.hxx>
#include <snprintf.h>
 
#include <algorithm>
#include <sstream>
#include <string>
 
#if defined(__GNUC__)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wshadow"
#pragma GCC diagnostic ignored "-Wunused-parameter"
#endif
#include <arrow/table.h>
#include <arrow/stl.h>
#if defined(__GNUC__)
#pragma GCC diagnostic pop
#endif
 
namespace ROOT {
namespace Internal {
namespace RDF {
 
// This is needed by Arrow 0.12.0 which dropped 
//
//      using ArrowType = ArrowType_;
//
// from ARROW_STL_CONVERSION
template <typename T>
struct RootConversionTraits {};
 
#define ROOT_ARROW_STL_CONVERSION(c_type, ArrowType_)  \
   template <>                                         \
   struct RootConversionTraits<c_type> {               \
   using ArrowType = ::arrow::ArrowType_;              \
   };
 
ROOT_ARROW_STL_CONVERSION(bool, BooleanType)
ROOT_ARROW_STL_CONVERSION(int8_t, Int8Type)
ROOT_ARROW_STL_CONVERSION(int16_t, Int16Type)
ROOT_ARROW_STL_CONVERSION(int32_t, Int32Type)
ROOT_ARROW_STL_CONVERSION(Long64_t, Int64Type)
ROOT_ARROW_STL_CONVERSION(uint8_t, UInt8Type)
ROOT_ARROW_STL_CONVERSION(uint16_t, UInt16Type)
ROOT_ARROW_STL_CONVERSION(uint32_t, UInt32Type)
ROOT_ARROW_STL_CONVERSION(ULong64_t, UInt64Type)
ROOT_ARROW_STL_CONVERSION(float, FloatType)
ROOT_ARROW_STL_CONVERSION(double, DoubleType)
ROOT_ARROW_STL_CONVERSION(std::string, StringType)
 
// Per slot visitor of an Array.
class ArrayPtrVisitor : public ::arrow::ArrayVisitor {
private:
   /// The pointer to update.
   void **fResult;
   bool fCachedBool{false}; // Booleans need to be unpacked, so we use a cached entry.
   // FIXME: I should really use a variant here
   RVec<float> fCachedRVecFloat;
   RVec<double> fCachedRVecDouble;
   RVec<ULong64_t> fCachedRVecULong64;
   RVec<UInt_t> fCachedRVecUInt;
   RVec<Long64_t> fCachedRVecLong64;
   RVec<Int_t> fCachedRVecInt;
   std::string fCachedString;
   /// The entry in the array which should be looked up.
   ULong64_t fCurrentEntry;
 
   template <typename T>
   void *getTypeErasedPtrFrom(arrow::ListArray const &array, int32_t entry, RVec<T> &cache)
   {
      using ArrowType = typename RootConversionTraits<T>::ArrowType;
      using ArrayType = typename arrow::TypeTraits<ArrowType>::ArrayType;
      auto values = reinterpret_cast<ArrayType *>(array.values().get());
      auto offset = array.value_offset(entry);
      // Here the cast to void* is a worksround while we figure out the
      // issues we have with long long types, signed and unsigned.
      RVec<T> tmp(reinterpret_cast<T *>((void *)values->raw_values()) + offset, array.value_length(entry));
      std::swap(cache, tmp);
      return (void *)(&cache);
   }
 
public:
   ArrayPtrVisitor(void **result) : fResult{result}, fCurrentEntry{0} {}
 
   void SetEntry(ULong64_t entry) { fCurrentEntry = entry; }
 
   /// Check if we are asking the same entry as before.
   virtual arrow::Status Visit(arrow::Int32Array const &array) final
   {
      *fResult = (void *)(array.raw_values() + fCurrentEntry);
      return arrow::Status::OK();
   }
 
   virtual arrow::Status Visit(arrow::Int64Array const &array) final
   {
      *fResult = (void *)(array.raw_values() + fCurrentEntry);
      return arrow::Status::OK();
   }
 
   /// Check if we are asking the same entry as before.
   virtual arrow::Status Visit(arrow::UInt32Array const &array) final
   {
      *fResult = (void *)(array.raw_values() + fCurrentEntry);
      return arrow::Status::OK();
   }
 
   virtual arrow::Status Visit(arrow::UInt64Array const &array) final
   {
      *fResult = (void *)(array.raw_values() + fCurrentEntry);
      return arrow::Status::OK();
   }
 
   virtual arrow::Status Visit(arrow::FloatArray const &array) final
   {
      *fResult = (void *)(array.raw_values() + fCurrentEntry);
      return arrow::Status::OK();
   }
 
   virtual arrow::Status Visit(arrow::DoubleArray const &array) final
   {
      *fResult = (void *)(array.raw_values() + fCurrentEntry);
      return arrow::Status::OK();
   }
 
   virtual arrow::Status Visit(arrow::BooleanArray const &array) final
   {
      fCachedBool = array.Value(fCurrentEntry);
      *fResult = reinterpret_cast<void *>(&fCachedBool);
      return arrow::Status::OK();
   }
 
   virtual arrow::Status Visit(arrow::StringArray const &array) final
   {
      fCachedString = array.GetString(fCurrentEntry);
      *fResult = reinterpret_cast<void *>(&fCachedString);
      return arrow::Status::OK();
   }
 
   virtual arrow::Status Visit(arrow::ListArray const &array) final
   {
      switch (array.value_type()->id()) {
      case arrow::Type::FLOAT: {
         *fResult = getTypeErasedPtrFrom(array, fCurrentEntry, fCachedRVecFloat);
         return arrow::Status::OK();
      }
      case arrow::Type::DOUBLE: {
         *fResult = getTypeErasedPtrFrom(array, fCurrentEntry, fCachedRVecDouble);
         return arrow::Status::OK();
      }
      case arrow::Type::UINT32: {
         *fResult = getTypeErasedPtrFrom(array, fCurrentEntry, fCachedRVecUInt);
         return arrow::Status::OK();
      }
      case arrow::Type::UINT64: {
         *fResult = getTypeErasedPtrFrom(array, fCurrentEntry, fCachedRVecULong64);
         return arrow::Status::OK();
      }
      case arrow::Type::INT32: {
         *fResult = getTypeErasedPtrFrom(array, fCurrentEntry, fCachedRVecInt);
         return arrow::Status::OK();
      }
      case arrow::Type::INT64: {
         *fResult = getTypeErasedPtrFrom(array, fCurrentEntry, fCachedRVecLong64);
         return arrow::Status::OK();
      }
      default: return arrow::Status::TypeError("Type not supported");
      }
   }
 
   using ::arrow::ArrayVisitor::Visit;
};
 
/// Helper class which keeps track for each slot where to get the entry.
class TValueGetter {
private:
   std::vector<void *> fValuesPtrPerSlot;
   std::vector<ULong64_t> fLastEntryPerSlot;
   std::vector<ULong64_t> fLastChunkPerSlot;
   std::vector<ULong64_t> fFirstEntryPerChunk;
   std::vector<ArrayPtrVisitor> fArrayVisitorPerSlot;
   /// Since data can be chunked in different arrays we need to construct an
   /// index which contains the first element of each chunk, so that we can
   /// quickly move to the correct chunk.
   std::vector<ULong64_t> fChunkIndex;
   arrow::ArrayVector fChunks;
 
public:
   TValueGetter(size_t slots, arrow::ArrayVector chunks)
      : fValuesPtrPerSlot(slots, nullptr), fLastEntryPerSlot(slots, 0), fLastChunkPerSlot(slots, 0), fChunks{chunks}
   {
      fChunkIndex.reserve(fChunks.size());
      size_t next = 0;
      for (auto &chunk : chunks) {
         fFirstEntryPerChunk.push_back(next);
         next += chunk->length();
         fChunkIndex.push_back(next);
      }
      for (size_t si = 0, se = fValuesPtrPerSlot.size(); si != se; ++si) {
         fArrayVisitorPerSlot.push_back(ArrayPtrVisitor{fValuesPtrPerSlot.data() + si});
      }
   }
 
   /// This returns the ptr to the ptr to actual data.
   std::vector<void *> SlotPtrs()
   {
      std::vector<void *> result;
      for (size_t i = 0; i < fValuesPtrPerSlot.size(); ++i) {
         result.push_back(fValuesPtrPerSlot.data() + i);
      }
      return result;
   }
 
   // Convenience method to avoid code duplication between
   // SetEntry and InitSlot
   void UncachedSlotLookup(unsigned int slot, ULong64_t entry)
   {
      // If entry is greater than the previous one,
      // we can skip all the chunks before the last one we
      // queried.
      size_t ci = 0;
      assert(slot < fLastChunkPerSlot.size());
      if (fLastEntryPerSlot[slot] < entry) {
         ci = fLastChunkPerSlot.at(slot);
      }
 
      for (size_t ce = fChunkIndex.size(); ci != ce; ++ci) {
         if (entry < fChunkIndex[ci]) {
            assert(slot < fLastChunkPerSlot.size());
            fLastChunkPerSlot[slot] = ci;
            break;
         }
      }
 
      // Update the pointer to the requested entry.
      // Notice that we need to find the entry
      auto chunk = fChunks.at(fLastChunkPerSlot[slot]);
      assert(slot < fArrayVisitorPerSlot.size());
      fArrayVisitorPerSlot[slot].SetEntry(entry - fFirstEntryPerChunk[fLastChunkPerSlot[slot]]);
      fLastEntryPerSlot[slot] = entry;
      auto status = chunk->Accept(fArrayVisitorPerSlot.data() + slot);
      if (!status.ok()) {
         std::string msg = "Could not get pointer for slot ";
         msg += std::to_string(slot) + " looking at entry " + std::to_string(entry);
         throw std::runtime_error(msg);
      }
   }
 
   /// Set the current entry to be retrieved
   void SetEntry(unsigned int slot, ULong64_t entry)
   {
      // Same entry as before
      if (fLastEntryPerSlot[slot] == entry) {
         return;
      }
      UncachedSlotLookup(slot, entry);
   }
};
 
} // namespace RDF
} // namespace Internal
 
namespace RDF {
 
/// Helper to get the contents of a given column
 
/// Helper to get the human readable name of type
class RDFTypeNameGetter : public ::arrow::TypeVisitor {
private:
   std::vector<std::string> fTypeName;
 
public:
   arrow::Status Visit(const arrow::Int64Type &) override
   {
      fTypeName.push_back("Long64_t");
      return arrow::Status::OK();
   }
   arrow::Status Visit(const arrow::Int32Type &) override
   {
      fTypeName.push_back("Int_t");
      return arrow::Status::OK();
   }
   arrow::Status Visit(const arrow::UInt64Type &) override
   {
      fTypeName.push_back("ULong64_t");
      return arrow::Status::OK();
   }
   arrow::Status Visit(const arrow::UInt32Type &) override
   {
      fTypeName.push_back("UInt_t");
      return arrow::Status::OK();
   }
   arrow::Status Visit(const arrow::FloatType &) override
   {
      fTypeName.push_back("float");
      return arrow::Status::OK();
   }
   arrow::Status Visit(const arrow::DoubleType &) override
   {
      fTypeName.push_back("double");
      return arrow::Status::OK();
   }
   arrow::Status Visit(const arrow::StringType &) override
   {
      fTypeName.push_back("string");
      return arrow::Status::OK();
   }
   arrow::Status Visit(const arrow::BooleanType &) override
   {
      fTypeName.push_back("bool");
      return arrow::Status::OK();
   }
   arrow::Status Visit(const arrow::ListType &l) override
   {
      /// Recursively visit List types and map them to
      /// an RVec. We accumulate the result of the recursion on
      /// fTypeName so that we can create the actual type
      /// when the recursion is done.
      fTypeName.push_back("ROOT::VecOps::RVec<%s>");
      return l.value_type()->Accept(this);
   }
   std::string result()
   {
      // This recursively builds a nested type.
      std::string result = "%s";
      char buffer[8192];
      for (size_t i = 0; i < fTypeName.size(); ++i) {
         snprintf(buffer, 8192, result.c_str(), fTypeName[i].c_str());
         result = buffer;
      }
      return result;
   }
 
   using ::arrow::TypeVisitor::Visit;
};
 
/// Helper to determine if a given Column is a supported type.
class VerifyValidColumnType : public ::arrow::TypeVisitor {
private:
public:
   virtual arrow::Status Visit(const arrow::Int64Type &) override { return arrow::Status::OK(); }
   virtual arrow::Status Visit(const arrow::UInt64Type &) override { return arrow::Status::OK(); }
   virtual arrow::Status Visit(const arrow::Int32Type &) override { return arrow::Status::OK(); }
   virtual arrow::Status Visit(const arrow::UInt32Type &) override { return arrow::Status::OK(); }
   virtual arrow::Status Visit(const arrow::FloatType &) override { return arrow::Status::OK(); }
   virtual arrow::Status Visit(const arrow::DoubleType &) override { return arrow::Status::OK(); }
   virtual arrow::Status Visit(const arrow::StringType &) override { return arrow::Status::OK(); }
   virtual arrow::Status Visit(const arrow::BooleanType &) override { return arrow::Status::OK(); }
   virtual arrow::Status Visit(const arrow::ListType &) override { return arrow::Status::OK(); }
 
   using ::arrow::TypeVisitor::Visit;
};
 
////////////////////////////////////////////////////////////////////////
/// Constructor to create an Arrow RDataSource for RDataFrame.
/// \param[in] inTable the arrow Table to observe.
/// \param[in] inColumns the name of the columns to use
/// In case columns is empty, we use all the columns found in the table
RArrowDS::RArrowDS(std::shared_ptr<arrow::Table> inTable, std::vector<std::string> const &inColumns)
   : fTable{inTable}, fColumnNames{inColumns}
{
   auto &columnNames = fColumnNames;
   auto &table = fTable;
   auto &index = fGetterIndex;
   // We want to allow people to specify which columns they
   // need so that we can think of upfront IO optimizations.
   auto filterWantedColumns = [&columnNames, &table]() {
      if (columnNames.empty()) {
         for (auto &field : table->schema()->fields()) {
            columnNames.push_back(field->name());
         }
      }
   };
 
   // To support both arrow 0.14.0 and 0.16.0
   using ColumnType = decltype(fTable->column(0));
 
   auto getRecordsFirstColumn = [&columnNames, &table]() {
      if (columnNames.empty()) {
         throw std::runtime_error("At least one column required");
      }
      const auto name = columnNames.front();
      const auto columnIdx = table->schema()->GetFieldIndex(name);
      return table->column(columnIdx)->length();
   };
 
   // All columns are supposed to have the same number of entries.
   auto verifyColumnSize = [&table](ColumnType column, int columnIdx, int nRecords) {
      if (column->length() != nRecords) {
         std::string msg = "Column ";
         msg += table->schema()->field(columnIdx)->name() + " has a different number of entries.";
         throw std::runtime_error(msg);
      }
   };
 
   /// For the moment we support only a few native types.
   auto verifyColumnType = [&table](ColumnType column, int columnIdx) {
      auto verifyType = std::make_unique<VerifyValidColumnType>();
      auto result = column->type()->Accept(verifyType.get());
      if (result.ok() == false) {
         std::string msg = "Column ";
         msg += table->schema()->field(columnIdx)->name() + " contains an unsupported type.";
         throw std::runtime_error(msg);
      }
   };
 
   /// This is used to create an index between the columnId
   /// and the associated getter.
   auto addColumnToGetterIndex = [&index](int columnId) { index.push_back(std::make_pair(columnId, index.size())); };
 
   /// Assuming we can get called more than once, we need to
   /// reset the getter index each time.
   auto resetGetterIndex = [&index]() { index.clear(); };
 
   /// This is what initialization actually does
   filterWantedColumns();
   resetGetterIndex();
   auto nRecords = getRecordsFirstColumn();
   for (auto &columnName : fColumnNames) {
      auto columnIdx = fTable->schema()->GetFieldIndex(columnName);
      addColumnToGetterIndex(columnIdx);
 
      auto column = fTable->column(columnIdx);
      verifyColumnSize(column, columnIdx, nRecords);
      verifyColumnType(column, columnIdx);
   }
}
 
////////////////////////////////////////////////////////////////////////
/// Destructor.
RArrowDS::~RArrowDS()
{
}
 
const std::vector<std::string> &RArrowDS::GetColumnNames() const
{
   return fColumnNames;
}
 
std::vector<std::pair<ULong64_t, ULong64_t>> RArrowDS::GetEntryRanges()
{
   auto entryRanges(std::move(fEntryRanges)); // empty fEntryRanges
   return entryRanges;
}
 
std::string RArrowDS::GetTypeName(std::string_view colName) const
{
   auto field = fTable->schema()->GetFieldByName(std::string(colName));
   if (!field) {
      std::string msg = "The dataset does not have column ";
      msg += colName;
      throw std::runtime_error(msg);
   }
   RDFTypeNameGetter typeGetter;
   auto status = field->type()->Accept(&typeGetter);
   if (status.ok() == false) {
      std::string msg = "RArrowDS does not support a column of type ";
      msg += field->type()->name();
      throw std::runtime_error(msg);
   }
   return typeGetter.result();
}
 
bool RArrowDS::HasColumn(std::string_view colName) const
{
   auto field = fTable->schema()->GetFieldByName(std::string(colName));
   if (!field) {
      return false;
   }
   return true;
}
 
bool RArrowDS::SetEntry(unsigned int slot, ULong64_t entry)
{
   for (auto link : fGetterIndex) {
      auto &getter = fValueGetters[link.second];
      getter->SetEntry(slot, entry);
   }
   return true;
}
 
void RArrowDS::InitSlot(unsigned int slot, ULong64_t entry)
{
   for (auto link : fGetterIndex) {
      auto &getter = fValueGetters[link.second];
      getter->UncachedSlotLookup(slot, entry);
   }
}
 
void splitInEqualRanges(std::vector<std::pair<ULong64_t, ULong64_t>> &ranges, int nRecords, unsigned int nSlots)
{
   ranges.clear();
   const auto chunkSize = nRecords / nSlots;
   const auto remainder = 1U == nSlots ? 0 : nRecords % nSlots;
   auto start = 0UL;
   auto end = 0UL;
   for (auto i : ROOT::TSeqU(nSlots)) {
      start = end;
      end += chunkSize;
      ranges.emplace_back(start, end);
      (void)i;
   }
   ranges.back().second += remainder;
}
 
int getNRecords(std::shared_ptr<arrow::Table> &table, std::vector<std::string> &columnNames)
{
   auto index = table->schema()->GetFieldIndex(columnNames.front());
   return table->column(index)->length();
};
 
template <typename T>
std::shared_ptr<arrow::ChunkedArray> getData(T p)
{
   return p->data();
}
 
template <>
std::shared_ptr<arrow::ChunkedArray>
getData<std::shared_ptr<arrow::ChunkedArray>>(std::shared_ptr<arrow::ChunkedArray> p)
{
   return p;
}
 
void RArrowDS::SetNSlots(unsigned int nSlots)
{
   assert(0U == fNSlots && "Setting the number of slots even if the number of slots is different from zero.");
   fNSlots = nSlots;
   // We dump all the previous getters structures and we rebuild it.
   auto nColumns = fGetterIndex.size();
 
   fValueGetters.clear();
   for (size_t ci = 0; ci != nColumns; ++ci) {
      auto chunkedArray = getData(fTable->column(fGetterIndex[ci].first));
      fValueGetters.emplace_back(std::make_unique<ROOT::Internal::RDF::TValueGetter>(nSlots, chunkedArray->chunks()));
   }
}
 
/// This needs to return a pointer to the pointer each value getter
/// will point to.
std::vector<void *> RArrowDS::GetColumnReadersImpl(std::string_view colName, const std::type_info &)
{
   auto &index = fGetterIndex;
   auto findGetterIndex = [&index](unsigned int column) {
      for (auto &entry : index) {
         if (entry.first == column) {
            return entry.second;
         }
      }
      throw std::runtime_error("No column found at index " + std::to_string(column));
   };
 
   const int columnIdx = fTable->schema()->GetFieldIndex(std::string(colName));
   const int getterIdx = findGetterIndex(columnIdx);
   assert(getterIdx != -1);
   assert((unsigned int)getterIdx < fValueGetters.size());
   return fValueGetters[getterIdx]->SlotPtrs();
}
 
void RArrowDS::Initialise()
{
   auto nRecords = getNRecords(fTable, fColumnNames);
   splitInEqualRanges(fEntryRanges, nRecords, fNSlots);
}
 
std::string RArrowDS::GetLabel()
{
   return "ArrowDS";
}
 
/// Creates a RDataFrame using an arrow::Table as input.
/// \param[in] table the arrow Table to observe.
/// \param[in] columnNames the name of the columns to use
/// In case columnNames is empty, we use all the columns found in the table
RDataFrame MakeArrowDataFrame(std::shared_ptr<arrow::Table> table, std::vector<std::string> const &columnNames)
{
   ROOT::RDataFrame tdf(std::make_unique<RArrowDS>(table, columnNames));
   return tdf;
}
 
} // namespace RDF
 
} // namespace ROOT