Rust is not interoperability with C++ for IPC schemas with dictionaries

[pierrebelzile]

In my application, I created in Rust RecordBatch's (arrow 12) that contain multiple dictionary fields. I serialized using streaming IPC and attempted to read with C++ (4.0 and 8.0). The server failed to recover the batches. It seems to get confused with associating the proper dictionary to the correct field. Seems to only see the last dictionary?

I should also point out that I attempted the opposite: Reading in Rust what was created in C++ and that also failed.

I reduced this to a batch with two dictionary fields and uploaded that to: https://github.com/pierrebelzile/arrow_ipc_test. This includes files with the binaries.

I'm hoping there is an expert IPC-person here that quickly knows what is wrong...

[viirya]

Hi @pierrebelzile , thanks for reporting the issue. I recently fixed some issues (e.g. #1636) on integrating with C++ for dictionaries. These fixes will be released soon in 14.0.0 (#1693). Can you re-check with 14.0.0 once it is released? Thank you.

[pierrebelzile]

Great. I will try as soon as it comes out. Any chance that you can test with your code if you are able to read the cpp_out.bin file in this repo?

[viirya]

Yea, I just downloaded cpp_out.bin file and ran a test with it. I can read the file by using IPC StreamReader:

let reader = StreamReader::try_new(file, None).unwrap();

reader.for_each(|batch| {
  let batch = batch.unwrap();
  println!("batch: {:?}", batch);
});

batch: RecordBatch { schema: Schema { fields: [Field { name: "as", data_type: Dictionary(Int8, Utf8), nullable: true, dict_id: 0, dict_is_ordered: false, metadata: None }, Field { name: "bs", data_type: Dictionary(Int8, Utf8), null
able: true, dict_id: 1, dict_is_ordered: false, metadata: None }], metadata: {} }, columns: [DictionaryArray {keys: PrimitiveArray<Int8>                                                                                               
[                                                                                                                  
  0,                                                     
  1,                                                                                                               
] values: StringArray                                    
[                                                                                                                  
  "a1",                                                  
  "a2",                                                                                                            
]}                                                       
, DictionaryArray {keys: PrimitiveArray<Int8>                                                                      
[                                                        
  0,                                                                                                               
  1,                                                     
] values: StringArray                                                                                              
[                                                        
  "b1",                                                                                                            
  "b2",                                                  
]}                                                                                                                 
], row_count: 2 }                                        

[viirya]

I think we can close this issue as it proves the file can be read now.

[pierrebelzile]

Closing given that almost certainly fixed. Will re-open if necessary.

[viirya]

Thank you @pierrebelzile. Yea, you can open this again if you find any issue on integrating this crate with C++.

[pierrebelzile]

I'm reopening because the problem is still there in the direction rust -> c++: An arrow IPC data with dictionaries written in Rust is not readable from Arrow C++ (version 8). I was able to confirm that the other direction (c++ -> rust) is working as expected with dictionaries.

[viirya]

@pierrebelzile Can you let me know how I can test the case (Rust -> C++)? Do you have reproducible example?

[pierrebelzile]

I updated the github repo to include instructions to build but I can't connect to it from my work to correctly finish setting up. In the meanwhile, I've single stepped the code in C++ and the issue is that the dictionary id seen in C++ is always 0 in cpp/src/arrow/ipc/reader.cc:786 (arrow 7.0). This causes the dictionaries to "overwrite" each other and that's why we only see one of them. I have no idea why it works in Rust...

[viirya]

Oh, that is because in your build_batch method, you create Field by:

fields.push(Field::new("as", array.data_type().clone(), true));
...
fields.push(Field::new("bs", array.data_type().clone(), true));

dict_id in Field is default to be 0.

So you duplicate dictionary ids.

I changed the code to the follows, and ran the C++ program without any issue:

fields.push(Field::new_dict("as", array.data_type().clone(), true, 0, false));
...
fields.push(Field::new_dict("bs", array.data_type().clone(), true, 1, false));

[pierrebelzile]

Thanks for finding that. I'm not used to this brecause in C++ the id is automatically assigned.

What will happen when (if not yet) the data type is a map<string dict key, string value>? How does one set the dict id then?

[viirya]

The dict_id is at Field. For Map, it has a Field which is a Struct field which has two fields for key and value. The dict id should be set at the nested field for the key.

[pierrebelzile]

Closing because issue was mine.