ParquetInPlaceEncoder.h

Go to the documentation of this file.

/*
 * Copyright 2022 HEAVY.AI, Inc.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#pragma once

#include "ParquetEncoder.h"
#include "ParquetMetadataValidator.h"
#include "ParquetShared.h"
#include "TypedParquetDetectBuffer.h"

#include <parquet/schema.h>
#include <parquet/types.h>

#include "Catalog/ColumnDescriptor.h"
#include "ForeignStorageBuffer.h"

namespace foreign_storage {

class ParquetInPlaceEncoder : public ParquetScalarEncoder {
 public:
  ParquetInPlaceEncoder(Data_Namespace::AbstractBuffer* buffer,
                        const size_t omnisci_data_type_byte_size,
                        const size_t parquet_data_type_byte_size)
      : ParquetScalarEncoder(buffer)
      , omnisci_data_type_byte_size_(omnisci_data_type_byte_size)
      , parquet_data_type_byte_size_(parquet_data_type_byte_size) {}

  virtual void reserve(const size_t num_elements) = 0;

  void appendData(const int16_t* def_levels,
                  const int16_t* rep_levels,
                  const int64_t values_read,
                  const int64_t levels_read,
                  int8_t* values) override {
    if (omnisci_data_type_byte_size_ < parquet_data_type_byte_size_) {
      for (int64_t i = 0; i < values_read; ++i) {
        encodeAndCopy(values + i * parquet_data_type_byte_size_,
                      values + i * omnisci_data_type_byte_size_);
      }
    }

    if (values_read < levels_read) {  // nulls exist
      decodeNullsAndEncodeData(
          values,
          def_levels,
          values_read,
          levels_read,
          omnisci_data_type_byte_size_ >= parquet_data_type_byte_size_);
    } else if (omnisci_data_type_byte_size_ >= parquet_data_type_byte_size_) {
      for (int64_t i = levels_read - 1; i >= 0; --i) {
        encodeAndCopy(values + i * parquet_data_type_byte_size_,
                      values + i * omnisci_data_type_byte_size_);
      }
    }

    buffer_->append(values, levels_read * omnisci_data_type_byte_size_);
  }

 protected:
  const size_t omnisci_data_type_byte_size_;
  const size_t parquet_data_type_byte_size_;

 private:
  void decodeNullsAndEncodeData(int8_t* data_ptr,
                                const int16_t* def_levels,
                                const int64_t values_read,
                                const int64_t levels_read,
                                const bool do_encoding) {
    for (int64_t i = levels_read - 1, j = values_read - 1; i >= 0; --i) {
      if (def_levels[i]) {  // not null
        CHECK(j >= 0);
        if (do_encoding) {
          encodeAndCopy(data_ptr + (j--) * parquet_data_type_byte_size_,
                        data_ptr + i * omnisci_data_type_byte_size_);
        } else {
          copy(data_ptr + (j--) * omnisci_data_type_byte_size_,
               data_ptr + i * omnisci_data_type_byte_size_);
        }
      } else {  // null
        setNull(data_ptr + i * omnisci_data_type_byte_size_);
      }
    }
  }
};

template <typename V, typename T, typename NullType = V>
class TypedParquetInPlaceEncoder : public ParquetInPlaceEncoder {
 public:
  TypedParquetInPlaceEncoder(Data_Namespace::AbstractBuffer* buffer,
                             const ColumnDescriptor* column_desciptor,
                             const parquet::ColumnDescriptor* parquet_column_descriptor)
      : ParquetInPlaceEncoder(
            buffer,
            sizeof(V),
            parquet::GetTypeByteSize(parquet_column_descriptor->physical_type()))
      , current_batch_offset_(0) {
    if (auto detect_buffer = dynamic_cast<TypedParquetDetectBuffer*>(buffer_)) {
      setDetectBufferConverterType();
    }
  }

  TypedParquetInPlaceEncoder(Data_Namespace::AbstractBuffer* buffer,
                             const size_t omnisci_data_type_byte_size,
                             const size_t parquet_data_type_byte_size)
      : ParquetInPlaceEncoder(buffer, sizeof(V), parquet_data_type_byte_size)
      , current_batch_offset_(0) {
    if (auto detect_buffer = dynamic_cast<TypedParquetDetectBuffer*>(buffer_)) {
      setDetectBufferConverterType();
    }
  }

  void validate(const int8_t* parquet_data,
                const int64_t j,
                const SQLTypeInfo& column_type) const override {
    // no-op by default
  }

  std::string integralTypeToString(const V& element) const {
    Datum d;
    d.bigintval = element;
    return DatumToString(d, ParquetEncoder::column_type_);
  }

  bool isIntegralType(const SQLTypeInfo& type) const {
    return type.is_timestamp() || type.is_time() || type.is_date() || type.is_boolean() ||
           type.is_decimal() || type.is_integer();
  }

  std::string elementToString(const V& element) const {
    // handle specialized cases that require specific formating when converting to string
    auto null_value = get_null_value<NullType>();
    if (element == null_value) {
      return "NULL";
    }
    if (isIntegralType(ParquetEncoder::column_type_)) {
      return integralTypeToString(element);
    }
    return std::to_string(element);
  }

  std::string encodedDataToString(const int8_t* bytes) const override {
    const auto& element = reinterpret_cast<const V*>(bytes)[0];
    return elementToString(element);
  }

  void setDetectBufferConverterType() {
    auto detect_buffer = dynamic_cast<TypedParquetDetectBuffer*>(buffer_);
    CHECK(detect_buffer);
    std::function<std::string(const V&)> element_to_string = [this](const V& element) {
      return this->elementToString(element);
    };
    detect_buffer->setConverterType<V>(element_to_string);
  }

  void validateUsingEncodersColumnType(const int8_t* parquet_data,
                                       const int64_t j) const override {
    validate(parquet_data, j, column_type_);
  }

  void reserve(const size_t num_append_elements) override {
    buffer_->reserve(buffer_->size() + (num_append_elements * sizeof(V)));
  }

  void appendDataTrackErrors(const int16_t* def_levels,
                             const int16_t* rep_levels,
                             const int64_t values_read,
                             const int64_t levels_read,
                             int8_t* values) override {
    CHECK(is_error_tracking_enabled_);
    int64_t i, j;
    for (i = 0, j = 0; i < levels_read; ++i) {
      if (def_levels[i]) {
        try {
          CHECK(j < values_read);
          validateUsingEncodersColumnType(values, j++);
        } catch (const std::runtime_error& error) {
          invalid_indices_.insert(current_chunk_offset_ + i);
        }
      } else if (column_type_.get_notnull()) {  // item is null for NOT NULL column
        invalid_indices_.insert(current_chunk_offset_ + i);
      }
    }
    current_chunk_offset_ += levels_read;
    appendData(def_levels, rep_levels, values_read, levels_read, values);
  }

  // TODO: this member largely mirrors `appendDataTrackErrors` but is only used
  // by the parquet-secific import FSI cut-over, and will be removed in the
  // future
  void validateAndAppendData(const int16_t* def_levels,
                             const int16_t* rep_levels,
                             const int64_t values_read,
                             const int64_t levels_read,
                             int8_t* values,
                             const SQLTypeInfo& column_type, /* may not be used */
                             InvalidRowGroupIndices& invalid_indices) override {
    int64_t i, j;
    for (i = 0, j = 0; i < levels_read; ++i) {
      if (def_levels[i]) {
        try {
          CHECK(j < values_read);
          validate(values, j++, column_type);
        } catch (const std::runtime_error& error) {
          invalid_indices.insert(current_batch_offset_ + i);
        }
      }
    }
    current_batch_offset_ += levels_read;
    appendData(def_levels, rep_levels, values_read, levels_read, values);
  }

  void eraseInvalidIndicesInBuffer(
      const InvalidRowGroupIndices& invalid_indices) override {
    if (invalid_indices.empty()) {
      return;
    }
    auto omnisci_data_values = reinterpret_cast<V*>(buffer_->getMemoryPtr());
    CHECK(buffer_->size() % omnisci_data_type_byte_size_ == 0);
    size_t num_elements = buffer_->size() / omnisci_data_type_byte_size_;
    std::remove_if(
        omnisci_data_values, omnisci_data_values + num_elements, [&](const V& value) {
          const V* start = omnisci_data_values;
          auto index = std::distance(start, &value);
          return invalid_indices.find(index) != invalid_indices.end();
        });
    size_t num_bytes_erased = invalid_indices.size() * omnisci_data_type_byte_size_;
    CHECK(num_bytes_erased <= buffer_->size());
    buffer_->setSize(buffer_->size() - num_bytes_erased);
  }

  void appendData(const int16_t* def_levels,
                  const int16_t* rep_levels,
                  const int64_t values_read,
                  const int64_t levels_read,
                  int8_t* values) override {
    if (std::is_same<V, T>::value && values_read == levels_read) {
      if (!encodingIsIdentityForSameTypes()) {
        for (int64_t i = 0; i < levels_read; ++i) {
          encodeAndCopy(values + i * omnisci_data_type_byte_size_,
                        values + i * omnisci_data_type_byte_size_);
        }
      }
      buffer_->append(values, levels_read * omnisci_data_type_byte_size_);
    } else {
      ParquetInPlaceEncoder::appendData(
          def_levels, rep_levels, values_read, levels_read, values);
    }
  }

  void encodeAndCopyContiguous(const int8_t* parquet_data_bytes,
                               int8_t* omnisci_data_bytes,
                               const size_t num_elements) override {
    auto parquet_data_ptr = reinterpret_cast<const T*>(parquet_data_bytes);
    auto omnisci_data_ptr = reinterpret_cast<V*>(omnisci_data_bytes);
    for (size_t i = 0; i < num_elements; ++i) {
      encodeAndCopy(reinterpret_cast<const int8_t*>(&parquet_data_ptr[i]),
                    reinterpret_cast<int8_t*>(&omnisci_data_ptr[i]));
    }
  }

  void setNull(int8_t* omnisci_data_bytes) override {
    auto& omnisci_data_value = reinterpret_cast<V*>(omnisci_data_bytes)[0];
    omnisci_data_value = get_null_value<NullType>();
  }

  void copy(const int8_t* omnisci_data_bytes_source,
            int8_t* omnisci_data_bytes_destination) override {
    const auto& omnisci_data_value_source =
        reinterpret_cast<const V*>(omnisci_data_bytes_source)[0];
    auto& omnisci_data_value_destination =
        reinterpret_cast<V*>(omnisci_data_bytes_destination)[0];
    omnisci_data_value_destination = omnisci_data_value_source;
  }

  std::shared_ptr<ChunkMetadata> getRowGroupMetadata(
      const parquet::RowGroupMetaData* group_metadata,
      const int parquet_column_index,
      const SQLTypeInfo& column_type) override {
    auto metadata = ParquetEncoder::createMetadata(column_type);
    auto column_metadata = group_metadata->ColumnChunk(parquet_column_index);

    // update statistics
    auto parquet_column_descriptor =
        group_metadata->schema()->Column(parquet_column_index);

    if (ParquetEncoder::validate_metadata_stats_ && group_metadata->num_rows() > 0) {
      auto stats = validate_and_get_column_metadata_statistics(column_metadata.get());
      if (stats->HasMinMax()) {
        // validate statistics if validation applicable as part of encoding
        if (auto parquet_scalar_validator =
                dynamic_cast<ParquetMetadataValidator*>(this)) {
          try {
            parquet_scalar_validator->validate(
                stats,
                column_type.is_array() ? column_type.get_elem_type() : column_type);
          } catch (const std::exception& e) {
            std::stringstream error_message;
            error_message
                << e.what() << " Error validating statistics of Parquet column '"
                << group_metadata->schema()->Column(parquet_column_index)->name() << "'";
            throw std::runtime_error(error_message.str());
          }
        }

        auto [stats_min, stats_max] = getEncodedStats(parquet_column_descriptor, stats);
        auto updated_chunk_stats = getUpdatedStats(stats_min, stats_max, column_type);
        metadata->fillChunkStats(updated_chunk_stats.min,
                                 updated_chunk_stats.max,
                                 metadata->chunkStats.has_nulls);
      }
      auto null_count = stats->null_count();
      validateNullCount(group_metadata->schema()->Column(parquet_column_index)->name(),
                        null_count,
                        column_type);
      metadata->chunkStats.has_nulls = null_count > 0;
    }

    // update sizing
    metadata->numBytes =
        sizeof(NullType)  // use NullType byte size since it is guaranteed to
                          // be the byte size of stored data
        * column_metadata->num_values();
    metadata->numElements = group_metadata->num_rows();

    return metadata;
  }

 protected:
  virtual bool encodingIsIdentityForSameTypes() const { return false; }

  std::pair<T, T> getUnencodedStats(std::shared_ptr<parquet::Statistics> stats) const {
    T stats_min = reinterpret_cast<T*>(stats->EncodeMin().data())[0];
    T stats_max = reinterpret_cast<T*>(stats->EncodeMax().data())[0];
    return {stats_min, stats_max};
  }

 private:
  static ChunkStats getUpdatedStats(V& stats_min,
                                    V& stats_max,
                                    const SQLTypeInfo& column_type) {
    ForeignStorageBuffer buffer;
    buffer.initEncoder(column_type);
    auto encoder = buffer.getEncoder();

    if (column_type.is_array()) {
      ArrayDatum min_datum(
          sizeof(V), reinterpret_cast<int8_t*>(&stats_min), false, DoNothingDeleter());
      ArrayDatum max_datum(
          sizeof(V), reinterpret_cast<int8_t*>(&stats_max), false, DoNothingDeleter());
      std::vector<ArrayDatum> min_max_datums{min_datum, max_datum};
      encoder->updateStats(&min_max_datums, 0, 1);
    } else {
      encoder->updateStats(reinterpret_cast<int8_t*>(&stats_min), 1);
      encoder->updateStats(reinterpret_cast<int8_t*>(&stats_max), 1);
    }
    auto updated_chunk_stats_metadata = std::make_shared<ChunkMetadata>();
    encoder->getMetadata(updated_chunk_stats_metadata);
    return updated_chunk_stats_metadata->chunkStats;
  }

  std::pair<V, V> getEncodedStats(
      const parquet::ColumnDescriptor* parquet_column_descriptor,
      std::shared_ptr<parquet::Statistics> stats) {
    V stats_min, stats_max;
    auto min_string = stats->EncodeMin();
    auto max_string = stats->EncodeMax();
    if constexpr (std::is_same<T, parquet::FixedLenByteArray>::value) {
      CHECK_EQ(parquet_column_descriptor->physical_type(),
               parquet::Type::FIXED_LEN_BYTE_ARRAY);
      parquet::FixedLenByteArray min_byte_array, max_byte_array;
      min_byte_array.ptr = reinterpret_cast<const uint8_t*>(min_string.data());
      max_byte_array.ptr = reinterpret_cast<const uint8_t*>(max_string.data());
      encodeAndCopy(reinterpret_cast<int8_t*>(&min_byte_array),
                    reinterpret_cast<int8_t*>(&stats_min));
      encodeAndCopy(reinterpret_cast<int8_t*>(&max_byte_array),
                    reinterpret_cast<int8_t*>(&stats_max));
    } else if constexpr (std::is_same<T, parquet::ByteArray>::value) {
      CHECK_EQ(parquet_column_descriptor->physical_type(), parquet::Type::BYTE_ARRAY);
      parquet::ByteArray min_byte_array, max_byte_array;
      min_byte_array.ptr = reinterpret_cast<const uint8_t*>(min_string.data());
      min_byte_array.len = min_string.length();
      max_byte_array.ptr = reinterpret_cast<const uint8_t*>(max_string.data());
      max_byte_array.len = max_string.length();
      encodeAndCopy(reinterpret_cast<int8_t*>(&min_byte_array),
                    reinterpret_cast<int8_t*>(&stats_min));
      encodeAndCopy(reinterpret_cast<int8_t*>(&max_byte_array),
                    reinterpret_cast<int8_t*>(&stats_max));
    } else {
      encodeAndCopy(reinterpret_cast<int8_t*>(min_string.data()),
                    reinterpret_cast<int8_t*>(&stats_min));
      encodeAndCopy(reinterpret_cast<int8_t*>(max_string.data()),
                    reinterpret_cast<int8_t*>(&stats_max));
    }
    return {stats_min, stats_max};
  }

  int64_t current_batch_offset_ = 0;
};

}  // namespace foreign_storage