[C++][Parquet] Getting only 0 when reading DELTA_BINARY_PACKED

[Anieway]

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

I have encountered a possible bug where when reading values one at a time with the low level API of parquet reader
(as done similarly in the example: /cpp/examples/parquet/low_level_api/reader_writer.cc)
in a row with DELTA_BINARY_PACKED encoding the results are all 0 regardless of file content.

aka this rows_read = int64_reader->ReadBatch(1, &definition_level, &repetition_level, &value, &values_read); gives a zero value
but this rows_read = int64_reader->ReadBatch(2, nullptr, nullptr, values, &values_read); gives the correct value.

So this seems not to occur when reading bigger batch-sizes (>1).

The problem might be somewhere within DeltaBitPackDecoder<DType>::GetInternal

Component(s)

Parquet

[rok]

Interesting! Are you building from source?

[mapleFU]

Hi, Anieway, I submit a fixing for encoder before: #14959

And I cannot reproduce this bug now, my code is running on master, and my running method is:

index 09af32289..19a96dc40 100644
--- a/cpp/examples/parquet/low_level_api/reader_writer.cc
+++ b/cpp/examples/parquet/low_level_api/reader_writer.cc
@@ -64,6 +64,8 @@ int main(int argc, char** argv) {
     // Add writer properties
     parquet::WriterProperties::Builder builder;
     builder.compression(parquet::Compression::SNAPPY);
+    builder.disable_dictionary();
+    builder.encoding("int32_field", parquet::Encoding::DELTA_BINARY_PACKED);
+    builder.encoding("int64_field", parquet::Encoding::DELTA_BINARY_PACKED);
     std::shared_ptr<parquet::WriterProperties> props = builder.build();

Can you pull the lastest code to check if the problem still here? If still a bug, can you provide the minimal code to reproduce this bug?

[tachyonwill]

I was able to reproduce. I believe the problem is

arrow/cpp/src/parquet/encoding.cc

Line 2482 in 463cdcc

if (ARROW_PREDICT_FALSE(i == max_values)) break;

when requesting a single value at a time, GetInternal will never actually call InitBlock. Thus the only thing that will ever be written to the output is the first value from the header. The line should probably be
if (ARROW_PREDICT_FALSE(i == total_value_count_)) break;

[mapleFU]

Seems that my previous trying is failed because I use RelWithDebInfo, and it skip assert. Thanks @tachyonwill
I submit a naive fix here: #15124

[mapleFU]

By the way, can anyone add parquet and C++ tag for this issue? I have no permission on this project and this issue :) @rok @Anieway