[PARQUET-1438] [C++] corrupted files produced on 32-bit architecture (i686)

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 1.12.0
• Assignee: Unassigned
• Reporter: Dmitry Kalinkin
• Created: 09/Oct/18 03:25
• Updated: 23/Jun/24 03:30
• Resolved: 17/Feb/19 08:09

Description

I'm using C++ API to convert some data to parquet files. I've noticed a regression when upgrading from arrow-cpp 0.10.0 + parquet-cpp 1.5.0 to arrow-cpp 0.11.0. The issue is that I can write parquet files without an error, but when I try to read those using pyarrow I get a segfault:

#0  0x00007fffd17c7f0f in int arrow::util::RleDecoder::GetBatchWithDictSpaced<float>(float const*, float*, int, int, unsigned char const*, long) ()
   from /nix/store/k6sy2ncjnkn5wnb2dq9m5f0qh446kjhg-arrow-cpp-0.11.0/lib/libparquet.so.11
#1  0x00007fffd17c8025 in parquet::DictionaryDecoder<parquet::DataType<(parquet::Type::type)4> >::DecodeSpaced(float*, int, int, unsigned char const*, long) ()
   from /nix/store/k6sy2ncjnkn5wnb2dq9m5f0qh446kjhg-arrow-cpp-0.11.0/lib/libparquet.so.11
#2  0x00007fffd17bcf0f in parquet::internal::TypedRecordReader<parquet::DataType<(parquet::Type::type)4> >::ReadRecordData(long) ()
   from /nix/store/k6sy2ncjnkn5wnb2dq9m5f0qh446kjhg-arrow-cpp-0.11.0/lib/libparquet.so.11
#3  0x00007fffd17bfbea in parquet::internal::TypedRecordReader<parquet::DataType<(parquet::Type::type)4> >::ReadRecords(long) ()
   from /nix/store/k6sy2ncjnkn5wnb2dq9m5f0qh446kjhg-arrow-cpp-0.11.0/lib/libparquet.so.11
#4  0x00007fffd179d2f7 in parquet::arrow::PrimitiveImpl::NextBatch(long, std::shared_ptr<arrow::Array>*) ()
   from /nix/store/k6sy2ncjnkn5wnb2dq9m5f0qh446kjhg-arrow-cpp-0.11.0/lib/libparquet.so.11
#5  0x00007fffd1797162 in parquet::arrow::ColumnReader::NextBatch(long, std::shared_ptr<arrow::Array>*) ()
   from /nix/store/k6sy2ncjnkn5wnb2dq9m5f0qh446kjhg-arrow-cpp-0.11.0/lib/libparquet.so.11
#6  0x00007fffd179a6e5 in parquet::arrow::FileReader::Impl::ReadSchemaField(int, std::vector<int, std::allocator<int> > const&, std::shared_ptr<arrow::Array>*) ()
   from /nix/store/k6sy2ncjnkn5wnb2dq9m5f0qh446kjhg-arrow-cpp-0.11.0/lib/libparquet.so.11
#7  0x00007fffd179aaad in parquet::arrow::FileReader::Impl::ReadTable(std::vector<int, std::allocator<int> > const&, std::shared_ptr<arrow::Table>*)::{lambda(int)#1}::operator()(int) const () from /nix/store/k6sy2ncjnkn5wnb2dq9m5f0qh446kjhg-arrow-cpp-0.11.0/lib/libparquet.so.11

I have not been able to dig to the bottom of the issue, but it seems like the problem reproduces only when I run 32 bit binaries. After I learned that, I found that 32 bit and 64 bit codes produce very different different parquet files for the same data. The sizes of the structures are clearly different if I look at their hexdumps. I'm attaching those example files. Reading "32.parquet" (produced using i686 binaries) will cause a segfault on macOS and linux, "64.parquet" will read just fine.