Skip to main content

Fury is a blazing fast multi-language serialization framework powered by jit, vectorization and zero-copy

Project description

Build Status Twitter

Fury is a blazing-fast multi-language serialization framework powered by jit(just-in-time compilation) and zero-copy, providing up to 170x performance and ultimate ease of use.


  • Multiple languages: Java/Python/C++/Golang/JavaScript/Rust/Scala/TypeScript.
  • Zero-copy: Cross-language out-of-band serialization inspired by pickle5 and off-heap read/write.
  • High performance: A highly-extensible JIT framework to generate serializer code at runtime in an async multi-thread way to speed serialization, providing 20-170x speed up by:
    • reduce memory access by inlining variables in generated code.
    • reduce virtual method invocation by inline call in generated code.
    • reduce conditional branching.
    • reduce hash lookup.
  • Multiple binary protocols: Object graph, row format, and so on.

In addition to cross-language serialization, Fury also features at:

  • Drop-in replace Java serialization frameworks such as JDK/Kryo/Hessian without modifying any code, but 100x faster. It can greatly improve the efficiency of high-performance RPC calls, data transfer, and object persistence.
  • 100% compatible with JDK serialization API with much faster implementation: supporting JDK writeObject/readObject/writeReplace/readResolve/readObjectNoData/Externalizable API.
  • Supports Java 8~21, Java 17+ record is supported too.
  • Support AOT compilation serialization for GraalVM native image, and no reflection/serialization json config are needed.
  • Supports shared and circular reference object serialization for golang.
  • Supports scala serialization
  • Supports automatic object serialization for golang.


Different scenarios have different serialization requirements. Fury designed and implemented multiple binary protocols for those requirements:

  • Cross-language object graph protocol:
    • Cross-language serialize any object automatically, no need for IDL definition, schema compilation and object to/from protocol conversion.
    • Support shared reference and circular reference, no duplicate data or recursion error.
    • Support object polymorphism.
  • Native java/python object graph protocol: Highly-optimized based on type system of the language.
  • Row format protocol: A cache-friendly binary random access format, supports skipping serialization and partial serialization, and can convert to column-format automatically.

New protocols can be easily added based on fury existing buffer, encoding, meta, codegen and other capabilities. All of those share the same codebase, and the optimization for one protocol can be reused by another protocol.


Different serialization frameworks are suitable for different scenarios, and benchmark results here are for reference only.

If you need to benchmark for your specific scenario, make sure all serialization frameworks are appropriately configured for that scenario.

Dynamic serialization frameworks support polymorphism and references, but they often come with a higher cost compared to static serialization frameworks, unless they utilize JIT techniques like Fury does. Because Fury generates code at runtime, it is recommended to warm up the system before collecting benchmark statistics.

Java Serialization

Title containing "compatible" represent schema compatible mode: support type forward/backward compatibility.

Title without "compatible" represent schema consistent mode: class schema must be the same between serialization and deserialization.

Struct is a class with 100 primitive fields, MediaContent is a class from jvm-serializers, Sample is a class from kryo benchmark.

See benchmarks for more benchmarks about type forward/backward compatibility, off-heap support, zero-copy serialization.



Nightly snapshot:

<!-- row/arrow format support -->
<!-- <dependency>
</dependency> -->

Release version:

<!-- row/arrow format support -->
<!-- <dependency>
</dependency> -->


libraryDependencies += "org.furyio" % "fury-core" % "0.4.1"


# Release version will be provided in the future.
pip install pyfury --pre


npm install @furyjs/fury


go get


Here we give a quick start about how to use fury, see user guide for more details about java, cross language, and row format.

Fury java object graph serialization

If you don't have cross-language requirements, using this mode will have better performance.

import io.fury.*;
import io.fury.config.*;
import java.util.*;

public class Example {
  public static void main(String[] args) {
    SomeClass object = new SomeClass();
    // Note that Fury instances should be reused between 
    // multiple serializations of different objects.
      Fury fury = Fury.builder().withLanguage(Language.JAVA)
        // Allow to deserialize objects unknown types, more flexible 
        // but may be insecure if the classes contains malicious code.
      // Registering types can reduce class name serialization overhead, but not mandatory.
      // If class registration enabled, all custom types must be registered.
      byte[] bytes = fury.serialize(object);
      ThreadSafeFury fury = Fury.builder().withLanguage(Language.JAVA)
        // Allow to deserialize objects unknown types, more flexible 
        // but may be insecure if the classes contains malicious code.
      byte[] bytes = fury.serialize(object);
      ThreadSafeFury fury = new ThreadLocalFury(classLoader -> {
        Fury f = Fury.builder().withLanguage(Language.JAVA)
        return f;
      byte[] bytes = fury.serialize(object);

Cross-language object graph serialization


import io.fury.*;
import io.fury.config.*;
import java.util.*;

public class ReferenceExample {
  public static class SomeClass {
    SomeClass f1;
    Map<String, String> f2;
    Map<String, String> f3;

  public static Object createObject() {
    SomeClass obj = new SomeClass();
    obj.f1 = obj;
    obj.f2 = ofHashMap("k1", "v1", "k2", "v2");
    obj.f3 = obj.f2;
    return obj;

  // mvn exec:java -Dexec.mainClass="io.fury.examples.ReferenceExample"
  public static void main(String[] args) {
    Fury fury = Fury.builder().withLanguage(Language.XLANG)
    fury.register(SomeClass.class, "example.SomeClass");
    byte[] bytes = fury.serialize(createObject());
    // bytes can be data serialized by other languages.


from typing import Dict
import pyfury

class SomeClass:
    f1: "SomeClass"
    f2: Dict[str, str]
    f3: Dict[str, str]

fury = pyfury.Fury(ref_tracking=True)
fury.register_class(SomeClass, type_tag="example.SomeClass")
obj = SomeClass()
obj.f2 = {"k1": "v1", "k2": "v2"}
obj.f1, obj.f3 = obj, obj.f2
data = fury.serialize(obj)
# bytes can be data serialized by other languages.


package main

import furygo ""
import "fmt"

func main() {
	type SomeClass struct {
		F1 *SomeClass
		F2 map[string]string
		F3 map[string]string
	fury := furygo.NewFury(true)
	if err := fury.RegisterTagType("example.SomeClass", SomeClass{}); err != nil {
	value := &SomeClass{F2: map[string]string{"k1": "v1", "k2": "v2"}}
	value.F3 = value.F2
	value.F1 = value
	bytes, err := fury.Marshal(value)
	if err != nil {
	var newValue interface{}
	// bytes can be data serialized by other languages.
	if err := fury.Unmarshal(bytes, &newValue); err != nil {

Row format


public class Bar {
  String f1;
  List<Long> f2;

public class Foo {
  int f1;
  List<Integer> f2;
  Map<String, Integer> f3;
  List<Bar> f4;

RowEncoder<Foo> encoder = Encoders.bean(Foo.class);
Foo foo = new Foo();
foo.f1 = 10;
foo.f2 = IntStream.range(0, 1000000).boxed().collect(Collectors.toList());
foo.f3 = IntStream.range(0, 1000000).boxed().collect(Collectors.toMap(i -> "k"+i, i->i));
List<Bar> bars = new ArrayList<>(1000000);
for (int i = 0; i < 1000000; i++) {
  Bar bar = new Bar();
  bar.f1 = "s"+i;
  bar.f2 = LongStream.range(0, 10).boxed().collect(Collectors.toList());
foo.f4 = bars;
// Can be zero-copy read by python
BinaryRow binaryRow = encoder.toRow(foo);
// can be data from python
Foo newFoo = encoder.fromRow(binaryRow);
// zero-copy read List<Integer> f2
BinaryArray binaryArray2 = binaryRow.getArray(1);
// zero-copy read List<Bar> f4
BinaryArray binaryArray4 = binaryRow.getArray(3);
// zero-copy read 11th element of `readList<Bar> f4`
BinaryRow barStruct = binaryArray4.getStruct(10);

// zero-copy read 6th of f2 of 11th element of `readList<Bar> f4`
RowEncoder<Bar> barEncoder = Encoders.bean(Bar.class);
// deserialize part of data.
Bar newBar = barEncoder.fromRow(barStruct);
Bar newBar2 = barEncoder.fromRow(binaryArray4.getStruct(20));


class Bar:
    f1: str
    f2: List[pa.int64]
class Foo:
    f1: pa.int32
    f2: List[pa.int32]
    f3: Dict[str, pa.int32]
    f4: List[Bar]

encoder = pyfury.encoder(Foo)
foo = Foo(f1=10, f2=list(range(1000_000)),
         f3={f"k{i}": i for i in range(1000_000)},
         f4=[Bar(f1=f"s{i}", f2=list(range(10))) for i in range(1000_000)])
binary: bytes = encoder.to_row(foo).to_bytes()
foo_row = pyfury.RowData(encoder.schema, binary)
print(foo_row.f2[100000], foo_row.f4[100000].f1, foo_row.f4[200000].f2[5])


Schema Compatibility

Fury java object graph serialization support class schema forward/backward compatibility. The serialization peer and deserialization peer can add/delete fields independently.

We plan to add support cross-language serialization after meta compression is finished.

Binary Compatibility

We are still improving our protocols, binary compatibility is not ensured between fury major releases for now. it's ensured between minor versions only. Please versioning your data by fury major version if you will upgrade fury in the future, see how to upgrade fury for further details.

Binary compatibility will be ensured when fury 1.0 is released.


Static serialization is secure. But dynamic serialization such as fury java/python native serialization supports deserializing unregistered types, which provides more dynamics and flexibility, but also introduce security risks.

For example, the deserialization may invoke init constructor or equals/hashCode method, if the method body contains malicious code, the system will be at risk.

Fury provides a class registration option that is enabled by default for such protocols, allowing only deserialization of trusted registered types or built-in types. Do not disable class registration unless you can ensure your environment is secure.

If this option is disabled, you are responsible for serialization security. You can configure io.fury.resolver.ClassChecker by ClassResolver#setClassChecker to control which classes are allowed for serialization.

How to Build

Please read the BUILD guide for instructions on how to build.

How to Contribute

Please read the CONTRIBUTING guide for instructions on how to contribute.

For ecosystem projects, please see


Licensed under the Apache License, Version 2.0

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distributions

pyfury-0.4.1-cp312-cp312-manylinux1_x86_64.whl (1.2 MB view hashes)

Uploaded CPython 3.12

pyfury-0.4.1-cp312-cp312-macosx_12_0_x86_64.whl (959.2 kB view hashes)

Uploaded CPython 3.12 macOS 12.0+ x86-64

pyfury-0.4.1-cp311-cp311-manylinux1_x86_64.whl (1.2 MB view hashes)

Uploaded CPython 3.11

pyfury-0.4.1-cp311-cp311-macosx_12_0_x86_64.whl (973.7 kB view hashes)

Uploaded CPython 3.11 macOS 12.0+ x86-64

pyfury-0.4.1-cp310-cp310-manylinux1_x86_64.whl (1.2 MB view hashes)

Uploaded CPython 3.10

pyfury-0.4.1-cp310-cp310-macosx_12_0_x86_64.whl (972.0 kB view hashes)

Uploaded CPython 3.10 macOS 12.0+ x86-64

pyfury-0.4.1-cp39-cp39-manylinux1_x86_64.whl (1.2 MB view hashes)

Uploaded CPython 3.9

pyfury-0.4.1-cp39-cp39-macosx_12_0_x86_64.whl (975.2 kB view hashes)

Uploaded CPython 3.9 macOS 12.0+ x86-64

pyfury-0.4.1-cp38-cp38-manylinux1_x86_64.whl (1.2 MB view hashes)

Uploaded CPython 3.8

pyfury-0.4.1-cp38-cp38-macosx_12_0_x86_64.whl (970.6 kB view hashes)

Uploaded CPython 3.8 macOS 12.0+ x86-64

pyfury-0.4.1-cp37-cp37m-manylinux1_x86_64.whl (1.2 MB view hashes)

Uploaded CPython 3.7m

pyfury-0.4.1-cp37-cp37m-macosx_12_0_x86_64.whl (952.2 kB view hashes)

Uploaded CPython 3.7m macOS 12.0+ x86-64

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page