apacheGH-47978: [C++][Parquet][CI] Add more compression codecs to fuzzing seed corpus (apache#47979)

pitrou · web-flow · commit 42f27abf56ac · 2025-10-29T17:31:21.000+01:00
### Rationale for this change 1. Add more compression codecs to seed corpus 2. Tweak fuzz target to make fuzzing slightly faster (around ~30% locally according to my measurements), which will allow testing more mutations per day ### Are these changes tested? Not specifically by CI, but hopefully they will make fuzzing more efficient. ### Are there any user-facing changes? No. * GitHub Issue: apache#47978 Authored-by: Antoine Pitrou <antoine@python.org> Signed-off-by: Antoine Pitrou <antoine@python.org>
diff --git a/cpp/src/parquet/arrow/generate_fuzz_corpus.cc b/cpp/src/parquet/arrow/generate_fuzz_corpus.cc
@@ -81,25 +81,68 @@ std::shared_ptr<Field> FieldForArray(const std::shared_ptr<Array>& array,
 }
 
 std::vector<WriteConfig> GetWriteConfigurations() {
+  auto default_properties_builder = [] {
+    auto builder = WriterProperties::Builder();
+    // Override current default of 1MB
+    builder.data_pagesize(10'000);
+    // Reduce max dictionary page size so that less pages are dict-encoded.
+    builder.dictionary_pagesize_limit(1'000);
+    // Emit various physical types for decimal columns
+    builder.enable_store_decimal_as_integer();
+    // DataPageV2 has more interesting features such as selective compression
+    builder.data_page_version(parquet::ParquetDataPageVersion::V2);
+    return builder;
+  };
+
+  auto default_arrow_properties_builder = [] {
+    auto builder = ArrowWriterProperties::Builder();
+    // Store the Arrow schema so as to exercise more data types when reading
+    builder.store_schema();
+    return builder;
+  };
+
   // clang-format off
-  auto w_brotli = WriterProperties::Builder()
-      .disable_dictionary("no_dict")
-      ->compression("compressed", Compression::BROTLI)
-      // Override current default of 1MB
-      ->data_pagesize(20'000)
-      // Reduce max dictionary page size so that less pages are dict-encoded.
-      ->dictionary_pagesize_limit(1'000)
-      // Emit various physical types for decimal columns
-      ->enable_store_decimal_as_integer()
+  auto w_uncompressed = default_properties_builder()
+      .build();
+  // compressed columns with dictionary disabled
+  auto w_brotli = default_properties_builder()
+      .disable_dictionary()
+      ->compression(Compression::BROTLI)
+      ->build();
+  auto w_gzip = default_properties_builder()
+      .disable_dictionary()
+      ->compression(Compression::GZIP)
       ->build();
-  // Store the Arrow schema so as to exercise more data types when reading
-  auto a_default = ArrowWriterProperties::Builder{}
-      .store_schema()
+  auto w_lz4 = default_properties_builder()
+      .disable_dictionary()
+      ->compression(Compression::LZ4)
       ->build();
+  auto w_snappy = default_properties_builder()
+      .disable_dictionary()
+      ->compression(Compression::SNAPPY)
+      ->build();
+  auto w_zstd = default_properties_builder()
+      .disable_dictionary()
+      ->compression(Compression::ZSTD)
+      ->build();
+  // v1 data pages
+  auto w_pages_v1 = default_properties_builder()
+      .disable_dictionary()
+      ->compression(Compression::LZ4)
+      ->data_page_version(parquet::ParquetDataPageVersion::V1)
+      ->build();
+
+  auto a_default = default_arrow_properties_builder().build();
   // clang-format on
 
   std::vector<WriteConfig> configs;
+  configs.push_back({w_uncompressed, a_default});
   configs.push_back({w_brotli, a_default});
+  configs.push_back({w_gzip, a_default});
+  configs.push_back({w_lz4, a_default});
+  configs.push_back({w_snappy, a_default});
+  configs.push_back({w_zstd, a_default});
+  configs.push_back({w_pages_v1, a_default});
   return configs;
 }
 
@@ -255,8 +298,6 @@ Result<std::vector<Column>> ExampleColumns(int32_t length,
 
   // TODO extension types: UUID, JSON, GEOMETRY, GEOGRAPHY
 
-  // A non-dict-encoded column (see GetWriteConfigurations)
-  columns.push_back({"no_dict", gen.String(length, 0, 30, null_probability)});
   // A column that should be quite compressible (see GetWriteConfigurations)
   columns.push_back({"compressed", gen.Int64(length, -10, 10, null_probability)});
 
diff --git a/cpp/src/parquet/arrow/reader.cc b/cpp/src/parquet/arrow/reader.cc
@@ -1414,21 +1414,32 @@ Status FuzzReader(std::unique_ptr<FileReader> reader) {
 }  // namespace
 
 Status FuzzReader(const uint8_t* data, int64_t size) {
-  auto buffer = std::make_shared<::arrow::Buffer>(data, size);
   Status st;
-  for (auto batch_size : std::vector<std::optional<int>>{std::nullopt, 1, 13, 300}) {
-    auto file = std::make_shared<::arrow::io::BufferReader>(buffer);
-    FileReaderBuilder builder;
+
+  auto buffer = std::make_shared<::arrow::Buffer>(data, size);
+  auto file = std::make_shared<::arrow::io::BufferReader>(buffer);
+  auto pool = ::arrow::default_memory_pool();
+
+  // Read Parquet file metadata only once, which will reduce iteration time slightly
+  std::shared_ptr<FileMetaData> pq_md;
+  BEGIN_PARQUET_CATCH_EXCEPTIONS
+  pq_md = ParquetFileReader::Open(file)->metadata();
+  END_PARQUET_CATCH_EXCEPTIONS
+
+  // Note that very small batch sizes probably make fuzzing slower
+  for (auto batch_size : std::vector<std::optional<int>>{std::nullopt, 13, 300}) {
     ArrowReaderProperties properties;
     if (batch_size) {
       properties.set_batch_size(batch_size.value());
     }
-    builder.properties(properties);
 
-    RETURN_NOT_OK(builder.Open(std::move(file)));
+    std::unique_ptr<ParquetFileReader> pq_file_reader;
+    BEGIN_PARQUET_CATCH_EXCEPTIONS
+    pq_file_reader = ParquetFileReader::Open(file, default_reader_properties(), pq_md);
+    END_PARQUET_CATCH_EXCEPTIONS
 
     std::unique_ptr<FileReader> reader;
-    RETURN_NOT_OK(builder.Build(&reader));
+    RETURN_NOT_OK(FileReader::Make(pool, std::move(pq_file_reader), properties, &reader));
     st &= FuzzReader(std::move(reader));
   }
   return st;
diff --git a/cpp/src/parquet/properties.h b/cpp/src/parquet/properties.h
@@ -324,8 +324,6 @@ class PARQUET_EXPORT WriterProperties {
           content_defined_chunking_options_(
               properties.content_defined_chunking_options()) {}
 
-    virtual ~Builder() {}
-
     /// \brief EXPERIMENTAL: Use content-defined page chunking for all columns.
     ///
     /// Optimize parquet files for content addressable storage (CAS) systems by writing
@@ -1198,7 +1196,6 @@ class PARQUET_EXPORT ArrowWriterProperties {
           use_threads_(kArrowDefaultUseThreads),
           executor_(NULLPTR),
           write_time_adjusted_to_utc_(false) {}
-    virtual ~Builder() = default;
 
     /// \brief Disable writing legacy int96 timestamps (default disabled).
     Builder* disable_deprecated_int96_timestamps() {
