sonic-rs

English | 中文

A fast Rust JSON library based on SIMD. It has some references to other open-source libraries like sonic_cpp, serde_json, sonic, simdjson, rust-std and more.

For Golang users to use sonic_rs, please see for_Golang_user.md

For users to migrate from serde_json to sonic_rs, can see serdejson_compatibility

Requirements/Notes

1. Faster in x86_64 or aarch64, other architecture is fallback and maybe very slower.

2. Requires Rust nightly version Support Stable Rust now.

3. Please add the compile options -C target-cpu=native

Quick to use sonic-rs

To ensure that SIMD instruction is used in sonic-rs, you need to add rustflags -C target-cpu=native and compile on the host machine. For example, Rust flags can be configured in Cargo config.

Add sonic-rs in Cargo.toml

[dependencies]
sonic-rs = "0.3"

Features

1. Serde into Rust struct as serde_json and serde.

2. Parse/Serialize JSON for untyped sonic_rs::Value, which can be mutable.

3. Get specific fields from a JSON with the blazing performance.

4. Use JSON as a lazy array or object iterator with the blazing performance.

5. Support LazyValue, Number and RawNumber(just like Golang's JsonNumber) in default.

6. The floating parsing precision is as Rust std in default.

Benchmark

The main optimization in sonic-rs is the use of SIMD. However, we do not use the two-stage SIMD algorithms from simd-json. We primarily use SIMD in the following scenarios:

1. parsing/serialize long JSON strings
2. parsing the fraction of float number
3. Getting a specific elem or field from JSON
4. Skipping white spaces when parsing JSON

More details about optimization can be found in performance.md.

Benchmarks environment:

Architecture:        x86_64
Model name:          Intel(R) Xeon(R) Platinum 8260 CPU @ 2.40GHz

AArch64 benchmark data can be found in benchmark_aarch64.md.

Benchmarks:

• Deserialize Struct: Deserialize the JSON into Rust struct. The defined struct and testdata is from json-benchmark

• Deseirlize Untyped: Deseialize the JSON into an untyped document

The serialize benchmarks work oppositely.

All deserialized benchmarks enabled UTF-8 validation and enabled float_roundtrip in serde-json to get sufficient precision as Rust std.

Deserialize Struct

The benchmark will parse JSON into a Rust struct, and there are no unknown fields in JSON text. All fields are parsed into struct fields in the JSON.

Sonic-rs is faster than simd-json because simd-json (Rust) first parses the JSON into a tape, then parses the tape into a Rust struct. Sonic-rs directly parses the JSON into a Rust struct, and there are no temporary data structures. The flamegraph is profiled in the citm_catalog case.

cargo bench --bench deserialize_struct -- --quiet

twitter/sonic_rs::from_slice_unchecked
                        time:   [694.74 µs 707.83 µs 723.19 µs]
twitter/sonic_rs::from_slice
                        time:   [796.44 µs 827.74 µs 861.30 µs]
twitter/simd_json::from_slice
                        time:   [1.0615 ms 1.0872 ms 1.1153 ms]
twitter/serde_json::from_slice
                        time:   [2.2659 ms 2.2895 ms 2.3167 ms]
twitter/serde_json::from_str
                        time:   [1.3504 ms 1.3842 ms 1.4246 ms]

citm_catalog/sonic_rs::from_slice_unchecked
                        time:   [1.2271 ms 1.2467 ms 1.2711 ms]
citm_catalog/sonic_rs::from_slice
                        time:   [1.3344 ms 1.3671 ms 1.4050 ms]
citm_catalog/simd_json::from_slice
                        time:   [2.0648 ms 2.0970 ms 2.1352 ms]
citm_catalog/serde_json::from_slice
                        time:   [2.9391 ms 2.9870 ms 3.0481 ms]
citm_catalog/serde_json::from_str
                        time:   [2.5736 ms 2.6079 ms 2.6518 ms]

canada/sonic_rs::from_slice_unchecked
                        time:   [3.7779 ms 3.8059 ms 3.8368 ms]
canada/sonic_rs::from_slice
                        time:   [3.9676 ms 4.0212 ms 4.0906 ms]
canada/simd_json::from_slice
                        time:   [7.9582 ms 8.0932 ms 8.2541 ms]
canada/serde_json::from_slice
                        time:   [9.2184 ms 9.3560 ms 9.5299 ms]
canada/serde_json::from_str
                        time:   [9.0383 ms 9.2563 ms 9.5048 ms]

Deserialize Untyped

The benchmark will parse JSON into a document. Sonic-rs seems faster for several reasons:

• There are also no temporary data structures in sonic-rs, as detailed above.
• Sonic-rs uses a memory arena for the whole document, resulting in fewer memory allocations, better cache-friendliness, and mutability.
• The JSON object in sonic_rs::Value is an array. Sonic-rs does not build a hashmap.

cargo bench --bench deserialize_value -- --quiet

twitter/sonic_rs_dom::from_slice
                        time:   [550.95 µs 556.10 µs 562.89 µs]
twitter/sonic_rs_dom::from_slice_unchecked
                        time:   [525.97 µs 530.26 µs 536.06 µs]
twitter/serde_json::from_slice
                        time:   [3.7599 ms 3.8009 ms 3.8513 ms]
twitter/serde_json::from_str
                        time:   [2.8618 ms 2.8960 ms 2.9396 ms]
twitter/simd_json::slice_to_owned_value
                        time:   [1.7302 ms 1.7557 ms 1.7881 ms]
twitter/simd_json::slice_to_borrowed_value
                        time:   [1.1870 ms 1.1951 ms 1.2039 ms]

canada/sonic_rs_dom::from_slice
                        time:   [4.9060 ms 4.9568 ms 5.0213 ms]
canada/sonic_rs_dom::from_slice_unchecked
                        time:   [4.7858 ms 4.8728 ms 4.9803 ms]
canada/serde_json::from_slice
                        time:   [16.689 ms 16.980 ms 17.335 ms]
canada/serde_json::from_str
                        time:   [16.398 ms 16.640 ms 16.932 ms]
canada/simd_json::slice_to_owned_value
                        time:   [12.627 ms 12.846 ms 13.070 ms]
canada/simd_json::slice_to_borrowed_value
                        time:   [12.030 ms 12.164 ms 12.323 ms]

citm_catalog/sonic_rs_dom::from_slice
                        time:   [1.6657 ms 1.6981 ms 1.7341 ms]
citm_catalog/sonic_rs_dom::from_slice_unchecked
                        time:   [1.5109 ms 1.5253 ms 1.5424 ms]
citm_catalog/serde_json::from_slice
                        time:   [8.1618 ms 8.2566 ms 8.3653 ms]
citm_catalog/serde_json::from_str
                        time:   [7.8652 ms 8.0706 ms 8.3074 ms]
citm_catalog/simd_json::slice_to_owned_value
                        time:   [3.9834 ms 4.0325 ms 4.0956 ms]
citm_catalog/simd_json::slice_to_borrowed_value
                        time:   [3.3196 ms 3.3433 ms 3.3689 ms]

Serialize Untyped

cargo bench --bench serialize_value -- --quiet

We serialize the document into a string. In the following benchmarks, sonic-rs appears faster for the twitter JSON. The twitter JSON contains many long JSON strings, which fit well with sonic-rs's SIMD optimization.

twitter/sonic_rs::to_string
                        time:   [380.90 µs 390.00 µs 400.38 µs]
twitter/serde_json::to_string
                        time:   [788.98 µs 797.34 µs 807.69 µs]
twitter/simd_json::to_string
                        time:   [965.66 µs 981.14 µs 998.08 µs]

citm_catalog/sonic_rs::to_string
                        time:   [805.85 µs 821.99 µs 841.06 µs]
citm_catalog/serde_json::to_string
                        time:   [1.8299 ms 1.8880 ms 1.9498 ms]
citm_catalog/simd_json::to_string
                        time:   [1.7356 ms 1.7636 ms 1.7972 ms]

canada/sonic_rs::to_string
                        time:   [6.5808 ms 6.7082 ms 6.8570 ms]
canada/serde_json::to_string
                        time:   [6.4800 ms 6.5747 ms 6.6893 ms]
canada/simd_json::to_string
                        time:   [7.3751 ms 7.5690 ms 7.7944 ms]

Serialize Struct

cargo bench --bench serialize_struct -- --quiet

The explanation is as mentioned above.

twitter/sonic_rs::to_string
                        time:   [434.03 µs 448.25 µs 463.97 µs]
twitter/simd_json::to_string
                        time:   [506.21 µs 515.54 µs 526.35 µs]
twitter/serde_json::to_string
                        time:   [719.70 µs 739.97 µs 762.69 µs]

canada/sonic_rs::to_string
                        time:   [4.6701 ms 4.7481 ms 4.8404 ms]
canada/simd_json::to_string
                        time:   [5.8072 ms 5.8793 ms 5.9625 ms]
canada/serde_json::to_string
                        time:   [4.5708 ms 4.6281 ms 4.6967 ms]

citm_catalog/sonic_rs::to_string
                        time:   [624.86 µs 629.54 µs 634.57 µs]
citm_catalog/simd_json::to_string
                        time:   [624.10 µs 633.55 µs 644.78 µs]
citm_catalog/serde_json::to_string
                        time:   [802.10 µs 814.15 µs 828.10 µs]

Get from JSON

cargo bench --bench get_from -- --quiet

The benchmark is getting a specific field from the twitter.json.

• sonic-rs::get_unchecked_from_str: without validate
• sonic-rs::get_from_str: with validate
• gjson::get_from_str: without validate

Sonic-rs utilize SIMD to quickly skip unnecessary fields in the unchecked case, thus enhancing the performance.

twitter/sonic-rs::get_unchecked_from_str
                        time:   [75.671 µs 76.766 µs 77.894 µs]
twitter/sonic-rs::get_from_str
                        time:   [430.45 µs 434.62 µs 439.43 µs]
twitter/gjson::get_from_str
                        time:   [359.61 µs 363.14 µs 367.19 µs]

Usage

Serde into Rust Type

Directly use the Deserialize or Serialize trait.

use sonic_rs::{Deserialize, Serialize}; 
// sonic-rs re-exported them from serde
// or use serde::{Deserialize, Serialize};

#[derive(Serialize, Deserialize)]
struct Person {
    name: String,
    age: u8,
    phones: Vec<String>,
}

fn main() {
    let data = r#"{
  "name": "Xiaoming",
  "age": 18,
  "phones": [
    "+123456"
  ]
}"#;
    let p: Person = sonic_rs::from_str(data).unwrap();
    assert_eq!(p.age, 18);
    assert_eq!(p.name, "Xiaoming");
    let out = sonic_rs::to_string_pretty(&p).unwrap();
    assert_eq!(out, data);
}

Get a field from JSON

Get a specific field from a JSON with the pointer path. The return is a LazyValue, which is a wrapper of a raw valid JSON slice.

We provide the get and get_unchecked apis. get_unchecked apis should be used in valid JSON, otherwise it may return unexpected result.

use sonic_rs::JsonValueTrait;
use sonic_rs::{get, get_unchecked, pointer};

fn main() {
    let path = pointer!["a", "b", "c", 1];
    let json = r#"
        {"u": 123, "a": {"b" : {"c": [null, "found"]}}}
    "#;
    let target = unsafe { get_unchecked(json, &path).unwrap() };
    assert_eq!(target.as_raw_str(), r#""found""#);
    assert_eq!(target.as_str().unwrap(), "found");

    let target = get(json, &path);
    assert_eq!(target.as_str().unwrap(), "found");
    assert_eq!(target.unwrap().as_raw_str(), r#""found""#);

    let path = pointer!["a", "b", "c", "d"];
    let json = r#"
        {"u": 123, "a": {"b" : {"c": [null, "found"]}}}
    "#;
    // not found from json
    let target = get(json, &path);
    assert!(target.is_err());
}

Parse and Serialize into untyped Value

Parse a JSON into a sonic_rs::Value.

use sonic_rs::{from_str, json};
use sonic_rs::JsonValueMutTrait;
use sonic_rs::{pointer, JsonValueTrait, Value};

fn main() {
    let json = r#"{
        "name": "Xiaoming",
        "obj": {},
        "arr": [],
        "age": 18,
        "address": {
            "city": "Beijing"
        },
        "phones": [
            "+123456"
        ]
    }"#;

    let mut root: Value = from_str(json).unwrap();

    // get key from value
    let age = root.get("age").as_i64();
    assert_eq!(age.unwrap_or_default(), 18);

    // get by index
    let first = root["phones"][0].as_str().unwrap();
    assert_eq!(first, "+123456");

    // get by pointer
    let phones = root.pointer(&pointer!["phones", 0]);
    assert_eq!(phones.as_str().unwrap(), "+123456");

    // convert to mutable object
    let obj = root.as_object_mut().unwrap();
    obj.insert(&"inserted", true);
    assert!(obj.contains_key(&"inserted"));

    let mut object = json!({ "A": 65, "B": 66, "C": 67 });
    *object.get_mut("A").unwrap() = json!({
        "code": 123,
        "success": false,
        "payload": {}
    });

    let mut val = json!(["A", "B", "C"]);
    *val.get_mut(2).unwrap() = json!("D");

    // serialize
    assert_eq!(serde_json::to_string(&val).unwrap(), r#"["A","B","D"]"#);
}

JSON Iterator

Parse an object or array JSON into a lazy iterator.

use bytes::Bytes;
use faststr::FastStr;
use sonic_rs::JsonValueTrait;
use sonic_rs::{to_array_iter, to_object_iter_unchecked};
fn main() {
    let json = Bytes::from(r#"[1, 2, 3, 4, 5, 6]"#);
    let iter = to_array_iter(&json);
    for (i, v) in iter.enumerate() {
        assert_eq!(i + 1, v.as_u64().unwrap() as usize);
    }

    let json = Bytes::from(r#"[1, 2, 3, 4, 5, 6"#);
    let iter = to_array_iter(&json);
    for elem in iter {
        // do something for each elem

        // deal with errors when invalid json
        if elem.is_err() {
            assert_eq!(
                elem.err().unwrap().to_string(),
                "Expected this character to be either a ',' or a ']' while parsing at line 1 column 17"
            );
        }
    }

    let json = FastStr::from(r#"{"a": null, "b":[1, 2, 3]}"#);
    let iter = unsafe { to_object_iter_unchecked(&json) };
    for ret in iter {
        // deal with errors
        if ret.is_err() {
            println!("{}", ret.unwrap_err());
            return;
        }

        let (k, v) = ret.unwrap();
        if k == "a" {
            assert!(v.is_null());
        } else if k == "b" {
            let iter = to_array_iter(v.as_raw_str());
            for (i, v) in iter.enumerate() {
                assert_eq!(i + 1, v.as_u64().unwrap() as usize);
            }
        }
    }
}

JSON LazyValue & Number & RawNumber

If we need to parse a JSON value as a raw string, we can use LazyValue.

If we need to parse a JSON number into an untyped type, we can use Number.

If we need to parse a JSON number without loss of precision, we can use RawNumber. It likes encoding/json.Number in Golang, and can also be parsed from a JSON string.

Detailed examples can be found in raw_value.rs and json_number.rs.

Error handle

Sonic's errors are followed as serde-json and have a display around the error position, examples in handle_error.rs.

FAQs

About UTF-8

By default, sonic-rs enable the UTF-8 validation, except for xx_unchecked APIs.

About floating point precision

By default, sonic-rs uses floating point precision consistent with the Rust standard library, and there is no need to add an extra float_roundtrip feature like serde-json to ensure floating point precision.

If you want to achieve lossless precision when parsing floating-point numbers, such as Golang encoding/json.Number and serde-json arbitrary_precision, you can use sonic_rs::RawNumber.

Acknowledgement

Thanks the following open-source libraries. sonic-rs has some references to other open-source libraries like sonic_cpp, serde_json, sonic, simdjson, yyjson, rust-std and so on.

We rewrote many SIMD algorithms from sonic-cpp/sonic/simdjson/yyjson for performance. We reused the de/ser codes and modified necessary parts from serde_json to make high compatibility with serde. We reused part codes about floating parsing from rust-std to make it more accurate.

Referenced papers:

1. Parsing Gigabytes of JSON per Second

2. JSONSki: streaming semi-structured data with bit-parallel fast-forwarding

Contributing

Please read CONTRIBUTING.md for information on contributing to sonic-rs.