pitrou commented on code in PR #37400:
URL: https://github.com/apache/arrow/pull/37400#discussion_r2689667662
##########
cpp/src/parquet/properties.h:
##########
@@ -169,6 +169,37 @@ static constexpr bool DEFAULT_IS_PAGE_INDEX_ENABLED = true;
static constexpr SizeStatisticsLevel DEFAULT_SIZE_STATISTICS_LEVEL =
SizeStatisticsLevel::PageAndColumnChunk;
+struct PARQUET_EXPORT BloomFilterOptions {
+ /// Expected number of distinct values (NDV) in the bloom filter.
+ ///
+ /// Bloom filters are most effective for high-cardinality columns. A good
default
+ /// is to set ndv equal to the number of rows. Lower values reduce disk
usage but
+ /// may not be worthwhile for very small NDVs.
+ ///
+ /// Increasing ndv (without increasing fpp) increases disk and memory usage.
+ int32_t ndv = 1 << 20;
+
+ /// False-positive probability (FPP) of the bloom filter.
+ ///
+ /// Lower FPP values require more disk and memory space. Recommended values
are
+ /// 0.1, 0.05, or 0.001. Very small values are counterproductive as the
bitset
+ /// may exceed the size of the actual data. Set ndv appropriately to minimize
+ /// space usage.
+ ///
+ /// Below is a table to demonstrate estimated size using common values.
+ ///
+ /// | ndv | fpp | bits/key | theoretical | actual (Po2) |
+ /// |:-----------|:------|:---------|:------------|:-------------|
+ /// | 100,000 | 0.10 | ~6.0 | 75 KB | **128 KB** |
+ /// | 100,000 | 0.01 | ~10.5 | 131 KB | **256 KB** |
Review Comment:
I'm not sure how you computed those numbers? Using the formula from
`OptimalNumOfBits` (`-8.0 / log(1 - pow(fpp, 1.0 / 8))` I get 5.77 bits / key
for fpp = 0.1 (and 9.68 bits / key for fpp = 0.01).
##########
cpp/src/parquet/bloom_filter_writer.cc:
##########
@@ -0,0 +1,264 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements. See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership. The ASF licenses this file
+// to you 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.
+
+#include "parquet/bloom_filter_writer.h"
+
+#include <map>
+#include <utility>
+
+#include "arrow/array.h"
+#include "arrow/io/interfaces.h"
+#include "arrow/type_traits.h"
+#include "arrow/util/bit_run_reader.h"
+#include "arrow/util/checked_cast.h"
+
+#include "parquet/exception.h"
+#include "parquet/metadata.h"
+#include "parquet/properties.h"
+#include "parquet/schema.h"
+#include "parquet/types.h"
+
+namespace parquet {
+
+constexpr int64_t kHashBatchSize = 256;
+
+template <typename ParquetType>
+TypedBloomFilterWriter<ParquetType>::TypedBloomFilterWriter(const
ColumnDescriptor* descr,
+ BloomFilter*
bloom_filter)
+ : descr_(descr), bloom_filter_(bloom_filter) {}
+
+template <typename ParquetType>
+void TypedBloomFilterWriter<ParquetType>::Update(const T* values, int64_t
num_values) {
+ ARROW_DCHECK(bloom_filter_ != nullptr);
+ std::array<uint64_t, kHashBatchSize> hashes;
+ for (int64_t i = 0; i < num_values; i += kHashBatchSize) {
+ auto batch_size = static_cast<int>(std::min(kHashBatchSize, num_values -
i));
+ if constexpr (std::is_same_v<ParquetType, FLBAType>) {
+ bloom_filter_->Hashes(values + i, descr_->type_length(), batch_size,
hashes.data());
+ } else {
+ bloom_filter_->Hashes(values + i, batch_size, hashes.data());
+ }
+ bloom_filter_->InsertHashes(hashes.data(), batch_size);
+ }
+}
+
+template <>
+void TypedBloomFilterWriter<BooleanType>::Update(const bool*, int64_t) {
+ throw ParquetException("Bloom filter is not supported for boolean type");
+}
+
+template <typename ParquetType>
+void TypedBloomFilterWriter<ParquetType>::UpdateSpaced(const T* values,
+ int64_t num_values,
+ const uint8_t*
valid_bits,
+ int64_t
valid_bits_offset) {
+ ARROW_DCHECK(bloom_filter_ != nullptr);
+ std::array<uint64_t, kHashBatchSize> hashes;
+ ::arrow::internal::VisitSetBitRunsVoid(
+ valid_bits, valid_bits_offset, num_values, [&](int64_t position, int64_t
length) {
+ for (int64_t i = 0; i < length; i += kHashBatchSize) {
+ auto batch_size = static_cast<int>(std::min(kHashBatchSize, length -
i));
+ if constexpr (std::is_same_v<ParquetType, FLBAType>) {
+ bloom_filter_->Hashes(values + i + position, descr_->type_length(),
+ batch_size, hashes.data());
+ } else {
+ bloom_filter_->Hashes(values + i + position, batch_size,
hashes.data());
+ }
+ bloom_filter_->InsertHashes(hashes.data(), batch_size);
+ }
+ });
+}
+
+template <>
+void TypedBloomFilterWriter<BooleanType>::UpdateSpaced(const bool*, int64_t,
+ const uint8_t*,
int64_t) {
+ throw ParquetException("Bloom filter is not supported for boolean type");
+}
+
+template <typename ParquetType>
+void TypedBloomFilterWriter<ParquetType>::Update(const ::arrow::Array& values)
{
+ ParquetException::NYI("Updating bloom filter is not implemented for array of
type: " +
+ values.type()->ToString());
+}
+
+namespace {
+
+template <typename ArrayType>
+void UpdateBinaryBloomFilter(BloomFilter& bloom_filter, const ArrayType&
array) {
+ std::array<ByteArray, kHashBatchSize> byte_arrays;
+ std::array<uint64_t, kHashBatchSize> hashes;
+ ::arrow::internal::VisitSetBitRunsVoid(
+ array.null_bitmap_data(), array.offset(), array.length(),
+ [&](int64_t position, int64_t length) {
+ for (int64_t i = 0; i < length; i += kHashBatchSize) {
+ auto batch_size = static_cast<int>(std::min(kHashBatchSize, length -
i));
+ for (int j = 0; j < batch_size; j++) {
+ byte_arrays[j] = array.GetView(position + i + j);
+ }
+ bloom_filter.Hashes(byte_arrays.data(), batch_size, hashes.data());
+ bloom_filter.InsertHashes(hashes.data(), batch_size);
+ }
+ });
+}
+
+} // namespace
+
+template <>
+void TypedBloomFilterWriter<ByteArrayType>::Update(const ::arrow::Array&
values) {
+ ARROW_DCHECK(bloom_filter_ != nullptr);
+ if (::arrow::is_binary_view_like(values.type_id())) {
+ UpdateBinaryBloomFilter(
+ *bloom_filter_,
+ ::arrow::internal::checked_cast<const
::arrow::BinaryViewArray&>(values));
+ } else if (::arrow::is_binary_like(values.type_id())) {
+ UpdateBinaryBloomFilter(
+ *bloom_filter_,
+ ::arrow::internal::checked_cast<const ::arrow::BinaryArray&>(values));
+ } else if (::arrow::is_large_binary_like(values.type_id())) {
+ UpdateBinaryBloomFilter(
+ *bloom_filter_,
+ ::arrow::internal::checked_cast<const
::arrow::LargeBinaryArray&>(values));
+ } else {
+ ParquetException::NYI("Bloom filter is not supported for this Arrow type:
" +
+ values.type()->ToString());
+ }
+}
+
+template class TypedBloomFilterWriter<BooleanType>;
+template class TypedBloomFilterWriter<Int32Type>;
+template class TypedBloomFilterWriter<Int64Type>;
+template class TypedBloomFilterWriter<Int96Type>;
+template class TypedBloomFilterWriter<FloatType>;
+template class TypedBloomFilterWriter<DoubleType>;
+template class TypedBloomFilterWriter<ByteArrayType>;
+template class TypedBloomFilterWriter<FLBAType>;
+
+namespace {
+
+/// \brief A concrete implementation of BloomFilterBuilder.
+///
+/// \note Column encryption for bloom filter is not implemented yet.
+class BloomFilterBuilderImpl : public BloomFilterBuilder {
+ public:
+ BloomFilterBuilderImpl(const SchemaDescriptor* schema,
+ const WriterProperties* properties)
+ : schema_(schema), properties_(properties) {}
+
+ void AppendRowGroup() override;
+
+ BloomFilter* CreateBloomFilter(int32_t column_ordinal) override;
+
+ IndexLocations WriteTo(::arrow::io::OutputStream* sink) override;
+
+ private:
+ /// Make sure column ordinal is not out of bound and the builder is in good
state.
+ void CheckState(int32_t column_ordinal) const {
+ if (finished_) {
+ throw ParquetException("BloomFilterBuilder is already finished.");
+ }
+ if (bloom_filters_.empty()) {
+ throw ParquetException("No row group appended to BloomFilterBuilder");
+ }
+ if (column_ordinal < 0 || column_ordinal >= schema_->num_columns()) {
+ throw ParquetException("Invalid column ordinal: " +
std::to_string(column_ordinal));
+ }
+ if (schema_->Column(column_ordinal)->physical_type() == Type::BOOLEAN) {
+ throw ParquetException("BloomFilterBuilder does not support boolean
type.");
+ }
+ }
+
+ const SchemaDescriptor* schema_;
+ const WriterProperties* properties_;
+ bool finished_ = false;
+
+ using RowGroupBloomFilters =
+ std::map</*column_id=*/int32_t, std::shared_ptr<BloomFilter>>;
+ std::vector<RowGroupBloomFilters> bloom_filters_; // indexed by row group
ordinal
+};
+
+void BloomFilterBuilderImpl::AppendRowGroup() {
+ if (finished_) {
+ throw ParquetException(
+ "Cannot append a new row group to a finished BloomFilterBuilder");
+ }
+ bloom_filters_.emplace_back();
+}
+
+BloomFilter* BloomFilterBuilderImpl::CreateBloomFilter(int32_t column_ordinal)
{
+ CheckState(column_ordinal);
+
+ auto opts =
properties_->bloom_filter_options(schema_->Column(column_ordinal)->path());
+ if (!opts.has_value()) {
+ return nullptr;
+ }
+
+ auto& curr_rg_bfs = *bloom_filters_.rbegin();
+ if (curr_rg_bfs.find(column_ordinal) != curr_rg_bfs.cend()) {
+ std::stringstream ss;
+ ss << "Bloom filter already exists for column: " << column_ordinal
+ << ", row group: " << (bloom_filters_.size() - 1);
+ throw ParquetException(ss.str());
+ }
+
+ auto bf =
std::make_unique<BlockSplitBloomFilter>(properties_->memory_pool());
+ bf->Init(BlockSplitBloomFilter::OptimalNumOfBytes(opts->ndv, opts->fpp));
+ return curr_rg_bfs.emplace(column_ordinal,
std::move(bf)).first->second.get();
+}
+
+IndexLocations BloomFilterBuilderImpl::WriteTo(::arrow::io::OutputStream*
sink) {
+ if (finished_) {
+ throw ParquetException("Cannot write a finished BloomFilterBuilder");
+ }
+ finished_ = true;
+
+ IndexLocations locations;
+
+ for (size_t i = 0; i != bloom_filters_.size(); ++i) {
+ auto& row_group_bloom_filters = bloom_filters_[i];
+ for (const auto& [column_id, filter] : row_group_bloom_filters) {
+ if (ARROW_PREDICT_FALSE(filter == nullptr)) {
+ throw ParquetException("Bloom filter cannot be null");
+ }
Review Comment:
Can this happen? We use `std::make_unique` which cannot return null.
##########
cpp/src/parquet/properties.h:
##########
@@ -169,6 +169,37 @@ static constexpr bool DEFAULT_IS_PAGE_INDEX_ENABLED = true;
static constexpr SizeStatisticsLevel DEFAULT_SIZE_STATISTICS_LEVEL =
SizeStatisticsLevel::PageAndColumnChunk;
+struct PARQUET_EXPORT BloomFilterOptions {
+ /// Expected number of distinct values (NDV) in the bloom filter.
+ ///
+ /// Bloom filters are most effective for high-cardinality columns. A good
default
+ /// is to set ndv equal to the number of rows. Lower values reduce disk
usage but
+ /// may not be worthwhile for very small NDVs.
+ ///
+ /// Increasing ndv (without increasing fpp) increases disk and memory usage.
+ int32_t ndv = 1 << 20;
+
+ /// False-positive probability (FPP) of the bloom filter.
+ ///
+ /// Lower FPP values require more disk and memory space. Recommended values
are
+ /// 0.1, 0.05, or 0.001. Very small values are counterproductive as the
bitset
+ /// may exceed the size of the actual data. Set ndv appropriately to minimize
+ /// space usage.
+ ///
+ /// Below is a table to demonstrate estimated size using common values.
+ ///
+ /// | ndv | fpp | bits/key | theoretical | actual (Po2) |
Review Comment:
"Po2" is cryptic, do you mean "power of 2" or something?
##########
cpp/src/parquet/properties.h:
##########
@@ -169,6 +169,37 @@ static constexpr bool DEFAULT_IS_PAGE_INDEX_ENABLED = true;
static constexpr SizeStatisticsLevel DEFAULT_SIZE_STATISTICS_LEVEL =
SizeStatisticsLevel::PageAndColumnChunk;
+struct PARQUET_EXPORT BloomFilterOptions {
+ /// Expected number of distinct values (NDV) in the bloom filter.
+ ///
+ /// Bloom filters are most effective for high-cardinality columns. A good
default
+ /// is to set ndv equal to the number of rows. Lower values reduce disk
usage but
+ /// may not be worthwhile for very small NDVs.
+ ///
+ /// Increasing ndv (without increasing fpp) increases disk and memory usage.
+ int32_t ndv = 1 << 20;
+
+ /// False-positive probability (FPP) of the bloom filter.
+ ///
+ /// Lower FPP values require more disk and memory space. Recommended values
are
Review Comment:
Perhaps say something about the growth law of disk/memory space, something
like "approximately proportional to log(1/fpp)"?
##########
cpp/src/parquet/properties.h:
##########
@@ -169,6 +169,37 @@ static constexpr bool DEFAULT_IS_PAGE_INDEX_ENABLED = true;
static constexpr SizeStatisticsLevel DEFAULT_SIZE_STATISTICS_LEVEL =
SizeStatisticsLevel::PageAndColumnChunk;
+struct PARQUET_EXPORT BloomFilterOptions {
+ /// Expected number of distinct values (NDV) in the bloom filter.
+ ///
+ /// Bloom filters are most effective for high-cardinality columns. A good
default
+ /// is to set ndv equal to the number of rows. Lower values reduce disk
usage but
+ /// may not be worthwhile for very small NDVs.
+ ///
+ /// Increasing ndv (without increasing fpp) increases disk and memory usage.
+ int32_t ndv = 1 << 20;
+
+ /// False-positive probability (FPP) of the bloom filter.
+ ///
+ /// Lower FPP values require more disk and memory space. Recommended values
are
+ /// 0.1, 0.05, or 0.001. Very small values are counterproductive as the
bitset
Review Comment:
Surely you mean 0.01 instead of 0.001?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
