Add support for reading columns as Apache Arrow arrays

[andygrove]

This is going to be easy. I have some ugly prototype code working already.

    let path = Path::new(&args[1]);
    let file = File::open(&path).unwrap();
    let parquet_reader = SerializedFileReader::new(file).unwrap();

    let row_group_reader = parquet_reader.get_row_group(0).unwrap();

    for i in 0..row_group_reader.num_columns() {
        match row_group_reader.get_column_reader(i) {
            Ok(ColumnReader::Int32ColumnReader(ref mut r)) => {

                let batch_size = 1024;
                let sz = mem::size_of::<i32>();
                let p = memory::allocate_aligned((batch_size * sz) as i64).unwrap();
                let ptr_i32 = unsafe { mem::transmute::<*const u8, *mut i32>(p) };
                let mut buf = unsafe {
                    slice::from_raw_parts_mut(ptr_i32, batch_size) };

//                let mut builder : Builder<i32> = Builder::with_capacity(1024);
//                let buffer = builder.finish();
                match r.read_batch(1024, None, None, &mut buf) {
                    Ok((count,_)) => {
                        let arrow_buffer = Buffer::from_raw_parts(ptr_i32, count as i32);
                        let arrow_array = Array::from(arrow_buffer);

                        match arrow_array.data() {
                            ArrayData::Int32(b) => {
                                println!("len: {}", b.len());
                                println!("data: {:?}", b.iter().collect::<Vec<i32>>());
                            },
                            _ => println!("wrong type")
                        }


                    },
                    _ => println!("error")
                }
            }
            _ => println!("column type not supported")
        }
    }

Now we need to come up with a real design and I probably need to add more helper methods to Arrow to make this easier.

[andygrove]

Here is a cleaned up version:

                let mut builder : Builder<i32> = Builder::with_capacity(batch_size);
                match r.read_batch(1024, None, None, builder.slice_mut(0, batch_size)) {
                    Ok((count,_)) => {
                        builder.set_len(count);
                        let arrow_array = Array::from(builder.finish());
                        match arrow_array.data() {
                            ArrayData::Int32(b) => {
                                println!("len: {}", b.len());
                                println!("data: {:?}", b.iter().collect::<Vec<i32>>());
                            },
                            _ => println!("wrong type")
                        }

[andygrove]

I've made good progress with integrating parquet-rs with datafusion .. I have examples like this working

    let mut ctx = ExecutionContext::local();

    let df = ctx.load_parquet("test/data/uk_cities.parquet/part-00000-bdf0c245-d300-4b28-bfdd-0f1f9cb898c4-c000.snappy.parquet").unwrap();

    ctx.register("uk_cities", df);

    // define the SQL statement
    let sql = "SELECT lat, lng FROM uk_cities";

    // create a data frame
    let df = ctx.sql(&sql).unwrap();

    df.show(10);

It only works for int32 and f32 columns though so far

[sunchao]

nice progress @andygrove ! really glad you can read parquet now using SQL.

[sadikovi]

Looks great! I am curious how Arrow handles/maps optional or repeated fields (when definition and/or repetition levels exist).

[andygrove]

I have merged the current Parquet support to master in DataFusion. I have added examples for both DataFrame and SQL API.

https://github.com/datafusion-rs/datafusion-rs/tree/master/examples

This is very rough code and I won't be promoting the fact that Parquet support is there until this is a little more complete and tested.

[liurenjie1024]

@andygrove @sunchao Are you guys working on this? I'm working on an implementation which takes the cpp version as a reference.

[sunchao]

@liurenjie1024 are you working on the arrow part or the parquet-rs part? some more work in the arrow repo needs to be done so that parquet-rs can read into arrow format . I'm working (slowly) on that part.

[liurenjie1024]

I'm working on the arrow part. Yes only part of cpp version can be implemented. I'll work out an early version.

Chao Sun [email protected] 于 2018年10月10日周三 下午5:27写道：

@liurenjie1024 https://github.com/liurenjie1024 are you working on the arrow https://github.com/apache/arrow part or the parquet-rs part? some more work in the arrow repo needs to be done so that parquet-rs can read into arrow format . I'm working (slowly) on that part.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/sunchao/parquet-rs/issues/79#issuecomment-428502989, or mute the thread https://github.com/notifications/unsubscribe-auth/ACpL5SeEZrsr5j8lEbv6H_y6s3s0SG-vks5ujb2OgaJpZM4TKCyA .

[nevi-me]

This is unrelated, but I've seen that CSV readers are being implemented in Arrow (I think Python, C++, and there's a Go one that's an open PR).

BurntSushi's rust-csv got me wondering whether it'd be possible to implement a native CSV to arrow reader, which I think would also extend nicely to parquet.

@sunchao @andygrove , do you know if there's been discussion of something like this? Also, would some codegen be required? It's something I'd love to try contribute to in the coming weeks.

[sunchao]

BurntSushi's rust-csv got me wondering whether it'd be possible to implement a native CSV to arrow reader, which I think would also extend nicely to parquet.

Yes it is certainly do-able - it needs to be implemented in the arrow repo though. Some pieces may still be missing and you're welcome to work on the arrow repo :)

Also I'm not sure how this extend to parquet. Can you elaborate?

[nevi-me]

Yes, I've been following the Rust impl in Arrow. When I'm ready, I'll ask about it in the mailing list before opening a JIRA (didn't see one).

The extension to parquet was more in concept than anything; in the sense that if I can read a csv to arrow, I'd be able to get csv -> parquet working. I understand that csv is a "simpler" format due to its flat nature.

Also, I haven't checked to see if datafusion-rs supports csv and parquet as sinks, so one'd be able to create table my_table<parquet> as select * from input<csv>. A lot of my curiousity comes as a result of having to watch paint dry at work while I use Spark, so I've been thinking of what I'd need to be able to do to replace some of my workflow with Rust.

[sunchao]

Yes, I've been following the Rust impl in Arrow. When I'm ready, I'll ask about it in the mailing list before opening a JIRA (didn't see one).

Yes feel free to create a JIRA :)

The extension to parquet was more in concept than anything; in the sense that if I can read a csv to arrow, I'd be able to get csv -> parquet working. I understand that csv is a "simpler" format due to its flat nature.

I see. Yes I think we discussed (cc @sadikovi ) multiple times about creating a CSV -> parquet converter. This will be convenient. We should certainly do it.

[liurenjie1024]

I'm working on an arrow reader implementation and have finished the first step, converting parquet schema to arrow schema in this PR, please help to review this.

[sunchao]

The umbrella task is #186. Please watch that issue for progress on this matter.