serde_arrow - convert sequences of Rust objects to Arrow arrays and back again

[Crate info] | [API docs] | Example | Related packages & performance | [How does it work?] | Status | License | Changes | Development

The arrow in-memory format is a powerful way to work with data frame like structures. The surrounding ecosystem includes a rich set of libraries, ranging from data frames such as Polars to query engines such as DataFusion. However, the API of the underlying Rust crates can be at times cumbersome to use due to the statically typed nature of Rust.

serde_arrow, offers a simple way to convert Rust objects into Arrow arrays and back. serde_arrow relies on the Serde package to interpret Rust objects. Therefore, adding support for serde_arrow to custom types is as easy as using Serde's derive macros.

In the Rust ecosystem there are two competing implementations of the arrow in-memory format. serde_arrow supports both arrow and arrow2 for schema tracing, serialization from Rust structs to arrays, and deserialization from arrays to Rust structs.

Example

Given a Rust structure

#[derive(Serialize, Deserialize)]
struct Record {
    a: f32,
    b: i32,
}

let records = vec![
    Record { a: 1.0, b: 1 },
    Record { a: 2.0, b: 2 },
    Record { a: 3.0, b: 3 },
];

Serialize to arrow RecordBatch

use arrow::datatypes::FieldRef;
use serde_arrow::schema::{TracingOptions, SerdeArrowSchema};

// Determine Arrow schema
let fields = Vec::<FieldRef>::from_type::<Record>(TracingOptions::default())?;

// Build a record batch
let batch = serde_arrow::to_record_batch(&fields, &records)?;

This RecordBatch can now be written to disk using ArrowWriter from the parquet crate.

let file = File::create("example.pq");
let mut writer = ArrowWriter::try_new(file, batch.schema(), None)?;
writer.write(&batch)?;
writer.close()?;

Serialize to arrow2 arrays

use arrow2::datatypes::Field;
use serde_arrow::schema::{TracingOptions, SerdeArrowSchema};

let fields = Vec::<Field>::from_type::<Record>(TracingOptions::default())?;
let arrays = serde_arrow::to_arrow2(&fields, &records)?;

These arrays can now be written to disk using the helper method defined in the arrow2 guide. For parquet:

use arrow2::{chunk::Chunk, datatypes::Schema};

// see https://jorgecarleitao.github.io/arrow2/io/parquet_write.html
write_chunk(
    "example.pq",
    Schema::from(fields),
    Chunk::new(arrays),
)?;

Usage from python

The written files can be read in Python via

# using polars
>>> import polars as pl
>>> pl.read_parquet("example.pq")
shape: (3, 2)
┌─────┬─────┐
│ a   ┆ b   │
│ --- ┆ --- │
│ f32 ┆ i32 │
╞═════╪═════╡
│ 1.0 ┆ 1   │
│ 2.0 ┆ 2   │
│ 3.0 ┆ 3   │
└─────┴─────┘

# using pandas
>>> import pandas as pd
>>> pd.read_parquet("example.pq")
     a  b
0  1.0  1
1  2.0  2
2  3.0  3

Related packages & Performance

• arrow: the JSON component of the official Arrow package supports serializing objects that support serialize via the Decoder object. It supports primitives types, structs and lists
• arrow2-convert: adds derive macros to convert objects from and to arrow2 arrays. It supports primitive types, structs, lists, and chrono's date time types. Enum support is experimental according to the Readme. If performance is the main objective, arrow2-convert is a good choice as it has no or minimal overhead over building the arrays manually.

The different implementation have the following performance differences, when compared to arrow2-convert:

The detailed runtimes of the benchmarks are listed below.

complex_common_serialize(100000)

label	time [ms]	arrow2_convert:	serde_arrow::to	serde_arrow::to	arrow_json::Rea
arrow2_convert::TryIntoArrow	51.73	1.00	0.32	0.28	0.13
serde_arrow::to_arrow2	160.55	3.10	1.00	0.88	0.41
serde_arrow::to_arrow	183.29	3.54	1.14	1.00	0.47
arrow_json::ReaderBuilder	391.25	7.56	2.44	2.13	1.00

complex_common_serialize(1000000)

label	time [ms]	arrow2_convert:	serde_arrow::to	serde_arrow::to	arrow_json::Rea
arrow2_convert::TryIntoArrow	509.11	1.00	0.33	0.26	0.14
serde_arrow::to_arrow2	1534.82	3.01	1.00	0.80	0.42
serde_arrow::to_arrow	1921.37	3.77	1.25	1.00	0.53
arrow_json::ReaderBuilder	3653.99	7.18	2.38	1.90	1.00

primitives_serialize(100000)

label	time [ms]	arrow2_convert:	serde_arrow::to	serde_arrow::to	arrow_json::Rea
arrow2_convert::TryIntoArrow	13.96	1.00	0.47	0.33	0.11
serde_arrow::to_arrow2	29.42	2.11	1.00	0.70	0.23
serde_arrow::to_arrow	41.94	3.01	1.43	1.00	0.33
arrow_json::ReaderBuilder	128.67	9.22	4.37	3.07	1.00

primitives_serialize(1000000)

label	time [ms]	arrow2_convert:	serde_arrow::to	serde_arrow::to	arrow_json::Rea
arrow2_convert::TryIntoArrow	133.19	1.00	0.47	0.31	0.09
serde_arrow::to_arrow2	284.80	2.14	1.00	0.67	0.20
serde_arrow::to_arrow	427.89	3.21	1.50	1.00	0.30
arrow_json::ReaderBuilder	1408.58	10.58	4.95	3.29	1.00

License

Copyright (c) 2021 - 2024 Christopher Prohm and contributors

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.