Skip to content

GH-50689: [C++][Parquet] Add IEEE-754 total order and nan count for floating types #50807

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
HuaHuaY wants to merge 1 commit into
base: main
Choose a base branch
from
Open

GH-50689: [C++][Parquet] Add IEEE-754 total order and nan count for floating types #50807

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions cpp/apidoc/Doxyfile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2487,6 +2487,7 @@ PREDEFINED = __attribute__(x)= \
ARROW_SUPPRESS_DEPRECATION_WARNING= \
ARROW_UNSUPPRESS_DEPRECATION_WARNING= \
GANDIVA_EXPORT= \
PARQUET_DEPRECATED(x)= \

@HuaHuaY HuaHuaY Aug 5, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Add this because I added a PARQUET_DEPRECATED at cpp/src/parquet/statistics.h and ci failed. https://github.com/apache/arrow/actions/runs/30975844655/job/92209544482

I believe this is a long-standing issue. If someone would like me to submit a separate PR to fix it, I can certainly do so.

PARQUET_EXPORT=

# If the MACRO_EXPANSION and EXPAND_ONLY_PREDEF tags are set to YES then this
Expand Down
40 changes: 33 additions & 7 deletions cpp/src/arrow/dataset/file_parquet.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,7 @@
#include "arrow/dataset/scanner.h"
#include "arrow/filesystem/path_util.h"
#include "arrow/table.h"
#include "arrow/type_traits.h"
#include "arrow/util/checked_cast.h"
#include "arrow/util/future.h"
#include "arrow/util/iterator.h"
Expand Down Expand Up @@ -370,7 +371,7 @@ std::optional<compute::Expression> ParquetFileFragment::EvaluateStatisticsAsExpr
const parquet::Statistics& statistics) {
auto field_expr = compute::field_ref(field_ref);

bool may_have_null = !statistics.HasNullCount() || statistics.null_count() > 0;
bool may_have_null = !statistics.HasNullCount() || statistics.null_count() != 0;
// Optimize for corner case where all values are nulls
if (statistics.num_values() == 0) {
// If there are no non-null values, column `field_ref` in the fragment
Expand All @@ -379,6 +380,31 @@ std::optional<compute::Expression> ParquetFileFragment::EvaluateStatisticsAsExpr
return is_null(std::move(field_expr));
}

auto with_null = [&](compute::Expression expression) {
if (may_have_null) {
return compute::or_(std::move(expression), is_null(field_expr));
}
return expression;
};
auto is_nan_expression = [&] { return compute::call("is_nan", {field_expr}); };

const bool is_floating_point = is_floating(field.type()->id());
const auto nan_count = statistics.nan_count();
const bool all_nan =
is_floating_point && nan_count.has_value() && *nan_count == statistics.num_values();
if (all_nan) {
return with_null(is_nan_expression());
}

if (field.type()->id() == Type::HALF_FLOAT) {
// TODO: Arrow compute has no HALF_FLOAT scalar comparison kernels, so numeric
// statistics expressions cannot be bound. GH-46858 tracks the scalar
// representation, while GH-50512 explains why HALF_FLOAT cannot simply be
// added to NumericTypes(). Statistics pruning is optional, so skip it instead
// of returning NotImplemented and failing the scan.
return std::nullopt;
}

std::shared_ptr<Scalar> min, max;
if (!StatisticsAsScalars(statistics, &min, &max).ok()) {
return std::nullopt;
Expand All @@ -393,10 +419,10 @@ std::optional<compute::Expression> ParquetFileFragment::EvaluateStatisticsAsExpr
if (min->Equals(*max)) {
auto single_value = compute::equal(field_expr, compute::literal(std::move(min)));

if (!may_have_null) {
return single_value;
if (is_floating_point && (!nan_count.has_value() || *nan_count != 0)) {

@pitrou pitrou Sep 2, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit: suggest creating a bool may_have_nan to make these if clauses clearer.

@pitrou pitrou Sep 2, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Or perhaps even a with_nan lambda just like with_null.

single_value = compute::or_(std::move(single_value), is_nan_expression());
}
return compute::or_(std::move(single_value), is_null(std::move(field_expr)));
return with_null(std::move(single_value));
}

auto lower_bound = compute::greater_equal(field_expr, compute::literal(min));
Expand All @@ -419,10 +445,10 @@ std::optional<compute::Expression> ParquetFileFragment::EvaluateStatisticsAsExpr
} else {
in_range = compute::and_(std::move(lower_bound), std::move(upper_bound));
}
if (may_have_null) {
return compute::or_(std::move(in_range), compute::is_null(std::move(field_expr)));
if (is_floating_point && (!nan_count.has_value() || *nan_count != 0)) {
in_range = compute::or_(std::move(in_range), is_nan_expression());
}
return in_range;
return with_null(std::move(in_range));
}
return std::nullopt;
}
Expand Down
92 changes: 91 additions & 1 deletion cpp/src/arrow/dataset/file_parquet_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,6 +38,7 @@
#include "arrow/testing/util.h"
#include "arrow/type.h"
#include "arrow/type_fwd.h"
#include "arrow/util/float16.h"
#include "arrow/util/io_util.h"
#include "arrow/util/logging_internal.h"
#include "arrow/util/range.h"
Expand Down Expand Up @@ -921,7 +922,7 @@ TEST(TestParquetStatistics, NullMax) {
auto statistics = reader->RowGroup(0)->metadata()->ColumnChunk(0)->statistics();
auto stat_expression =
ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *statistics);
EXPECT_EQ(stat_expression->ToString(), "(x >= 1)");
EXPECT_EQ(stat_expression->ToString(), "((x >= 1) or is_nan(x))");
}

TEST(TestParquetStatistics, NoNullCount) {
Expand Down Expand Up @@ -974,6 +975,95 @@ TEST(TestParquetStatistics, NoNullCount) {
}
}

template <typename T>
void TestNaNCount(const std::shared_ptr<DataType>& type,
const ::parquet::schema::NodePtr& parquet_node) {
auto field = ::arrow::field("x", type);
auto dataset_schema = ::arrow::schema({field});
::parquet::ColumnDescriptor descr(parquet_node, 0, 0);
Comment thread
wgtmac marked this conversation as resolved.
auto encode = [](T value) {
return std::string(reinterpret_cast<const char*>(&value), sizeof(value));
};
auto check_expression = [&](const std::optional<compute::Expression>& expression,
const char* expected) {
ASSERT_TRUE(expression.has_value());
EXPECT_EQ(expected, expression->ToString());
ASSERT_OK(expression->Bind(*dataset_schema));
};

::parquet::EncodedStatistics encoded_stats;
encoded_stats.set_min(encode(T{1})).set_max(encode(T{100})).set_null_count(0);
auto stats = ::parquet::Statistics::Make(&descr, &encoded_stats, 10);
auto expression = ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *stats);
check_expression(expression, "(((x >= 1) and (x <= 100)) or is_nan(x))");

encoded_stats.set_nan_count(0);
stats = ::parquet::Statistics::Make(&descr, &encoded_stats, 10);
expression = ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *stats);
check_expression(expression, "((x >= 1) and (x <= 100))");

encoded_stats.set_nan_count(2);
stats = ::parquet::Statistics::Make(&descr, &encoded_stats, 10);
expression = ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *stats);
check_expression(expression, "(((x >= 1) and (x <= 100)) or is_nan(x))");

encoded_stats.set_null_count(1);
stats = ::parquet::Statistics::Make(&descr, &encoded_stats, 10);
expression = ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *stats);
check_expression(expression,
"((((x >= 1) and (x <= 100)) or is_nan(x)) or "
"is_null(x, {nan_is_null=false}))");

encoded_stats.set_null_count(0);
encoded_stats.ClearMinMax();
stats = ::parquet::Statistics::Make(&descr, &encoded_stats, 2);
expression = ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *stats);
check_expression(expression, "is_nan(x)");
}

TEST(TestParquetStatistics, NaNCount) {
TestNaNCount<float>(float32(),
::parquet::schema::Float("x", ::parquet::Repetition::REQUIRED));
TestNaNCount<double>(float64(),
::parquet::schema::Double("x", ::parquet::Repetition::REQUIRED));
}

TEST(TestParquetStatistics, HalfFloatNaNCount) {
auto field = ::arrow::field("x", float16());
auto parquet_node = ::parquet::schema::PrimitiveNode::Make(
"x", ::parquet::Repetition::REQUIRED, ::parquet::LogicalType::Float16(),
::parquet::Type::FIXED_LEN_BYTE_ARRAY, 2);
::parquet::ColumnDescriptor descr(parquet_node, 0, 0);
auto encode = [](util::Float16 value) {
const auto bytes = value.ToLittleEndian();
return std::string(reinterpret_cast<const char*>(bytes.data()), bytes.size());
};

::parquet::EncodedStatistics encoded_stats;
encoded_stats.set_min(encode(util::Float16(-1.0f)))
.set_max(encode(util::Float16(1.0f)))
.set_null_count(0)
.set_nan_count(1);
auto stats = ::parquet::Statistics::Make(&descr, &encoded_stats, 3);
auto expression = ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *stats);
ASSERT_FALSE(expression.has_value());

encoded_stats.ClearMinMax();
encoded_stats.set_nan_count(3);
stats = ::parquet::Statistics::Make(&descr, &encoded_stats, 3);
expression = ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *stats);
ASSERT_TRUE(expression.has_value());
EXPECT_EQ("is_nan(x)", expression->ToString());
ASSERT_OK(expression->Bind(*::arrow::schema({field})));

encoded_stats.set_null_count(1);
stats = ::parquet::Statistics::Make(&descr, &encoded_stats, 3);
expression = ParquetFileFragment::EvaluateStatisticsAsExpression(*field, *stats);
ASSERT_TRUE(expression.has_value());
EXPECT_EQ("(is_nan(x) or is_null(x, {nan_is_null=false}))", expression->ToString());
ASSERT_OK(expression->Bind(*::arrow::schema({field})));
}

TEST_F(TestParquetFileFormat, MultithreadedScanRegression) {
// GH-38438: This test is similar to MultithreadedScan, but it try to use self
// designed Executor and DelayedBufferReader to mock async execution to make
Expand Down
51 changes: 51 additions & 0 deletions cpp/src/parquet/arrow/arrow_reader_writer_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,8 @@
#include "gmock/gmock.h"
#include "gtest/gtest.h"

#include <algorithm>
#include <array>
#include <cstdint>
#include <functional>
#include <set>
Expand Down Expand Up @@ -59,6 +61,7 @@
#include "arrow/util/key_value_metadata.h"
#include "arrow/util/logging_internal.h"
#include "arrow/util/range.h"
#include "arrow/util/ubsan.h"

#ifdef ARROW_CSV
# include "arrow/csv/api.h"
Expand All @@ -75,6 +78,7 @@
#include "parquet/arrow/writer.h"
#include "parquet/column_writer.h"
#include "parquet/file_writer.h"
#include "parquet/page_index.h"
#include "parquet/properties.h"
#include "parquet/test_util.h"
#include "parquet/types.h"
Expand Down Expand Up @@ -3672,6 +3676,53 @@ TEST(TestArrowReadWrite, NonUniqueDictionaryValues) {
}
}

TEST(TestArrowReadWrite, FloatingDictionaryBits) {
// Float32: -sNaN(payload=1), +qNaN(payload=2), +0, -0, 1.0f.
const std::array<uint32_t, 5> dictionary_bits{0xff800001, 0x7fc00002, 0x00000000,
0x80000000, 0x3f800000};
std::vector<float> dictionary_values(dictionary_bits.size());
std::transform(dictionary_bits.begin(), dictionary_bits.end(),
dictionary_values.begin(),
[](uint32_t bits) { return ::arrow::util::SafeCopy<float>(bits); });
std::shared_ptr<Array> dictionary;
::arrow::ArrayFromVector<::arrow::FloatType>(dictionary_values, &dictionary);
auto indices = ArrayFromJSON(::arrow::int32(), "[0, 0, 1, 2, 3, 4]");
ASSERT_OK_AND_ASSIGN(auto values, DictionaryArray::FromArrays(indices, dictionary));
auto table =
Table::Make(::arrow::schema({::arrow::field("values", values->type())}), {values});

auto properties = WriterProperties::Builder().enable_write_page_index()->build();
ASSERT_OK_AND_ASSIGN(auto buffer,
WriteTableToBuffer(table, table->num_rows(), properties));
auto parquet_reader = ParquetFileReader::Open(std::make_shared<BufferReader>(buffer));
auto metadata = parquet_reader->metadata();
ASSERT_EQ(ColumnOrder::IEEE_754_TOTAL_ORDER,
metadata->schema()->Column(0)->column_order().get_order());
auto statistics = metadata->RowGroup(0)->ColumnChunk(0)->statistics();
ASSERT_EQ(std::make_optional<int64_t>(3), statistics->nan_count());

auto column_index =
parquet_reader->GetPageIndexReader()->RowGroup(0)->GetColumnIndex(0);
ASSERT_NE(nullptr, column_index);
ASSERT_TRUE(column_index->has_nan_counts());
EXPECT_THAT(column_index->nan_counts(), ::testing::ElementsAre(3));

std::unique_ptr<FileReader> arrow_reader;
FileReaderBuilder builder;
ASSERT_OK(builder.Open(std::make_shared<BufferReader>(buffer)));
ASSERT_OK(builder.Build(&arrow_reader));
ASSERT_OK_AND_ASSIGN(auto read_table, arrow_reader->ReadTable());
auto actual =
checked_pointer_cast<::arrow::FloatArray>(read_table->column(0)->chunk(0));
const std::array<uint32_t, 6> expected_bits{dictionary_bits[0], dictionary_bits[0],
dictionary_bits[1], dictionary_bits[2],
dictionary_bits[3], dictionary_bits[4]};
for (int64_t value_index = 0; value_index < actual->length(); ++value_index) {
ASSERT_EQ(expected_bits[value_index],
::arrow::util::SafeCopy<uint32_t>(actual->Value(value_index)));
}
}

TEST(TestArrowReadWrite, DictionaryIndexBitwidthRoundtrip) {
// GH-30302: the bitwidth of Arrow dictionary indices should be preserved
for (const auto& index_type :
Expand Down
Loading
Loading