[PARQUET-1882] [C++] Writing an all-null column and then reading it with buffered_stream aborts the process

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: cpp-4.0.0
• Component/s: parquet-cpp
• Labels: pull-request-available

Environment

Windows 10 64-bit, MSVC

• Assignee: Micah Kornfield
• Reporter: Eric Gorelik
• Created: 09/Jul/20 16:10
• Updated: 12/Jul/20 21:14
• Resolved: 12/Jul/20 21:14

Description

When writing a column unbuffered that contains only nulls, a 0-byte dictionary page gets written. When then reading the resulting file with buffered_stream enabled, the column reader gets the length of the page (which is 0), and then tries to read that many bytes from the underlying input stream.

parquet/column_reader.cc, SerializedPageReader::NextPage

 

int compressed_len = current_page_header_.compressed_page_size;
int uncompressed_len = current_page_header_.uncompressed_page_size;

// Read the compressed data page.
std::shared_ptr<Buffer> page_buffer;
PARQUET_THROW_NOT_OK(stream_->Read(compressed_len, &page_buffer));

 

BufferedInputStream::Read, however, has an assertion that the bytes to read is strictly positive, so the assertion fails and aborts the process.

arrow/io/buffered.cc, BufferedInputStream::Impl

 

Status Read(int64_t nbytes, int64_t* bytes_read, void* out) {        
  ARROW_CHECK_GT(nbytes, 0);

 

 

Issue Links

• links to: GitHub Pull Request #7718