[Go][Parquet] Trouble using the C++ reader to read a Parquet file written with the Go writer

[tschaub]

Describe the bug, including details regarding any error messages, version, and platform.

Version: 7ef517e31ec3
OS: macOS 13 arm64

I'm uncertain if this is user error, an issue with the Go packages, or an issue with the C++ reader. I've put together a test that demonstrates the issue here: https://github.com/tschaub/parquet-issue-38503

I'm trying to use the pqarrow package to read an input Parquet file, transform some of the data, and write an output Parquet file. In the linked test case, there is no transformation step. So the test uses a pqarrow.FileReader, gets a pqarrow.RowGroupReader for each row group, reads each column as an arrow.Chunked, and uses a pqarrow.ArrowColumnWriter to write out the same.

When I try to use the C++ parquet-reader to read in the output file, I see the following error:

# parquet-reader output.parquet > /dev/null
Parquet error: Malformed levels. min: 2 max: 2 out of range.  Max Level: 1

This same test passes for other Parquet files. I originally encountered the problem with one of the Overture Maps Parquet files, and the linked test case is based on a subset of that data using only two columns and a single row.

Summarizing

file	C++ reader	Go reader
input.parquet	✅	✅
output.parquet	❌	✅

Component(s)

Go, Parquet

[mapleFU]

Sorry for late reply. Run case in https://github.com/tschaub/parquet-issue-38503 can generate the case?

[mapleFU]

# parquet-reader output.parquet > /dev/null
Parquet error: Malformed levels. min: 2 max: 2 out of range.  Max Level: 1

This is weird 😅 This means go generate level == 2 when max-level == 1. Let me checkout the reason.

[mapleFU]

Updated: I think C++ reader checks max-def-level, and it's 1. (and get 2). So it report the error.

The problem is that go writer has a bug here. I'll find out and fix it.

[mapleFU]

After rethink the impl, I found it's the impl's problem rather than a bug.

	ctx := context.Background()
	numFields := len(arrowReader.Manifest.Fields)
	numRowGroups := fileReader.NumRowGroups()
	for rowGroupIndex := 0; rowGroupIndex < numRowGroups; rowGroupIndex += 1 {
		rowGroupReader := arrowReader.RowGroup(rowGroupIndex)
		rowGroupWriter := fileWriter.AppendRowGroup()
		for fieldNum := 0; fieldNum < numFields; fieldNum += 1 {
			arr, err := rowGroupReader.Column(fieldNum).Read(ctx)
			fmt.Println("Read field: ", fieldNum)
			if err != nil {
				return nil, err
			}
			colWriter, err := pqarrow.NewArrowColumnWriter(arr, 0, int64(arr.Len()), arrowReader.Manifest, rowGroupWriter, fieldNum)
			if err != nil {
				return nil, err
			}
			if err := colWriter.Write(ctx); err != nil {
				return nil, err
			}
		}
	}
	if err := fileWriter.Close(); err != nil {
		return nil, err
	}

colWriter, err := pqarrow.NewArrowColumnWriter(arr, 0, int64(arr.Len()), arrowReader.Manifest, rowGroupWriter, fieldNum)

The code above is abit dangerous, the real code should be:

	numRowGroups := fileReader.NumRowGroups()
	for rowGroupIndex := 0; rowGroupIndex < numRowGroups; rowGroupIndex += 1 {
		rowGroupReader := arrowReader.RowGroup(rowGroupIndex)
		rowGroupWriter := fileWriter.AppendRowGroup()
		colIdx := 0
		for fieldNum := 0; fieldNum < numFields; fieldNum += 1 {
			arr, err := rowGroupReader.Column(fieldNum).Read(ctx)
			fmt.Println("Read field: ", fieldNum)
			for _, v := range arr.Chunks() {
				fmt.Println("CHUNK:" , v)
			}
			if err != nil {
				return nil, err
			}
			colWriter, err := pqarrow.NewArrowColumnWriter(arr, 0, int64(arr.Len()), arrowReader.Manifest, rowGroupWriter, colIdx)
			if err != nil {
				return nil, err
			}
			if err := colWriter.Write(ctx); err != nil {
				return nil, err
			}
			colIdx += colWriter.LeafCount()
		}
	}

colWriter.LeafCount() cannot be get now, maybe we need using other way.

[mapleFU]

And currently, the generated file is a bad file here...(in your example)