[IMPALA-9831] TestScannersFuzzing::test_fuzz_alltypes() hits DCHECK in parquet-page-reader.cc

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 4.0.0
• Fix Version/s: Impala 4.0.0
• Component/s: Backend
• Labels: flaky
• Target Version: Impala 4.0.0
• Epic Color: ghx-label-9
• Assignee: Joe McDonnell
• Reporter: Joe McDonnell
• Created: 05/Jun/20 16:50
• Updated: 17/Jun/20 15:04
• Resolved: 17/Jun/20 15:04

Description

In a recent precommit job, an Impalad crashed with the following DCHECK:

F0604 01:18:36.921769 30923 parquet-page-reader.cc:67] b64df3da7eea7c65:16f9c6e800000001] Check failed: col_end < file_desc.file_length (6820 vs. 6820) 

The assert is checking that the end of a column is before the end of the file. This must be true, because the footer takes up space at the end of the file.

The code for this DCHECK is:

  int64_t col_end = col_start + col_len;
  // Already validated in ValidateColumnOffsets()
  DCHECK_GT(col_end, 0);
  DCHECK_LT(col_end, file_desc.file_length); <---------

This mentions that this was already validated in ParquetMetadataUtils::ValidateColumnOffsets(). That is where the problem is:

int64_t col_len = col_chunk.meta_data.total_compressed_size;
int64_t col_end = col_start + col_len;
if (col_end <= 0 || col_end > file_length) {
  return Status(Substitute("Parquet file '$0': metadata is corrupt. Column $1 has "
      "invalid column offsets (offset=$2, size=$3, file_size=$4).", filename, i,
      col_start, col_len, file_length));
}

The condition should be "col_end >= file_length".

If we knew the size of the parquet footer, this check could be stricter as well.