I wish we would focus on making tooling better for W3C EXI (Binary XML encoding) instead of inventing new formats. Just being fast isn't enough, I don't see many using Aeron/SBT, it need a ecosystem - which XML does have.
Binary XML encoding (like W3C EXI) is useful in some contexts, but it’s generally not as efficient as modern binary serialization formats. It also can’t naturally express shared or circular reference semantics, which are important for complex object graphs.
Fory’s format was designed from the ground up to handle those cases efficiently, while still enabling cross‑language compatibility and schema evolution.
I am not sure if W3C EXI, or ASN.1 BER or something else is better, but agree that using DOP (rather than OOP) design principles is the right answer -- which means focusing on the encoding first, and working backwards towards the languages / clients.
DOP is great, but there’s always a gap between DOP and OOP. That gap is where Fory comes in. Right now, Fory takes an OOP‑first approach, but next we’ll add a DOP path by introducing an optional IDL — bridging the two styles. My goal is for the IDL to also support optional OOP‑style expressiveness, so teams can choose the balance that fits their needs.
DOP is very interesting, I like this idea too — most DOP approaches are implemented via an IDL, which is another valid direction. I plan to support that in Fory. I want to give users the freedom to choose the model that works best for them.
It seems if the serialization object is not a "Fory" struct, then it is forced to go through to/from conversion as part of the measured serialization work:
I'd think that the to/from Fory types is shouldn't be part of the tests.
Also, when used in an actual system tonic would be providing a 8KB buffer to write into, not just a Vec::default() that may need to be resized multiple times:
I can see the source of an 10x improvement on an Intel(R) Xeon(R) Gold 6136 CPU @ 3.00GHz, but it drops to 3x improvement when I remove the to/from that clones or collects Vecs, and always allocate an 8K Vec instead of a ::Default for the writable buffer.
If anything, the benches should be updated in a tower service / codec generics style where other formats like protobuf do not use any Fory-related code at all.
Note also that Fory has some writer pool that is utilized during the tests:
Benchmarking ecommerce_data/fory_serialize/medium: Collecting 100 samples in estimated 5.0494 s (197k it
ecommerce_data/fory_serialize/medium
time: [25.373 µs 25.605 µs 25.916 µs]
change: [-2.0973% -0.9263% +0.2852%] (p = 0.15 > 0.05)
No change in performance detected.
Found 4 outliers among 100 measurements (4.00%)
2 (2.00%) high mild
2 (2.00%) high severe
Compared to original bench for Protobuf/Prost:
Benchmarking ecommerce_data/protobuf_serialize/medium: Collecting 100 samples in estimated 5.0419 s (20k
ecommerce_data/protobuf_serialize/medium
time: [248.85 µs 251.04 µs 253.86 µs]
Found 18 outliers among 100 measurements (18.00%)
8 (8.00%) high mild
10 (10.00%) high severe
However after allocating 8K instead of ::Default and removing to/from it for an updated protobuf bench:
fair_ecommerce_data/protobuf_serialize/medium
time: [73.114 µs 73.885 µs 74.911 µs]
change: [-1.8410% -0.6702% +0.5190%] (p = 0.30 > 0.05)
No change in performance detected.
Found 14 outliers among 100 measurements (14.00%)
2 (2.00%) high mild
12 (12.00%) high severe
The Rust benchmarks in Fory are intended more as end‑to‑end benchmarks for typical OOP‑style application scenarios, not just raw buffer write speed.
Protobuf is very much a DOP (data‑oriented programming) approach — which is great for some systems. But in many complex applications, especially those using polymorphism, teams don’t want to couple Protobuf‑generated message structs directly into their domain models. Generated types are harder to extend, and if you embed them everywhere (fields, parameters, return types), switching to another serialization framework later becomes almost impossible without touching huge parts of the codebase.
In large systems, it’s common to define independent domain model structs used throughout the codebase, and only convert to/from the Protobuf messages at the serialization boundary. That conversion step is exactly what’s represented in our benchmarks — because it’s what happens in many real deployments.
There’s also the type‑system gap: for example, if your Rust struct has a Box<dyn Trait> field, representing that cleanly in Protobuf is tricky. You might fall back to a oneof, but that essentially generates an enum variant, which often isn’t what users actually want for polymorphic behavior.
So, yes — we include the conversion in our measurements intentionally, to reflect the real‑world large systems practices.
Yes, I agree that protos usually should only be used at the serialization boundary, as well as the slightly off-topic idea that the generated code should be private to the package and/or binary.
So to reflect the real‑world practices, the benchmark code should then allocate and give the protobuf serializer an 8K Vec like in tonic, and not an empty one that may require multiple re-allocations?
DOP is great for certain scenarios, but there’s always a gap between DOP and OOP. That gap is where an extra domain model and the conversion step come in — especially in systems that rely heavily on polymorphism or want to keep serialization types decoupled from core business models.
Regarding design tradeoffs: I am very skeptical that this can be made to work for the long run in a cross-language way without formalizing the on-the-wire contract via IDL or similar.
In my experience, while starting from a language to arrive at the serialization often feels more ergonomic (e.g. RPC style) in the start, it hides too much of what's going on from the users and over time suffers greatly from programming language / runtime changes - the latter multiplied by the number of languages or frameworks supported.
That’s a fair point — with more languages in the mix, having a formal schema can definitely help prevent drift.
The way I think about it is:
• Single‑language projects often work best without an IDL — it keeps things simple and avoids extra steps.
• Two languages – both IDL and no‑IDL approaches can work, depending on the team’s habits.
• Three or more – an IDL can be really useful as a single source of truth and to avoid manually writing struct definitions in every language.
For Apache Fory, my plan is to add optional IDL support, so teams who want that “single truth” can generate definitions automatically, and others can continue with language‑first development. My hope is to give people flexibility to choose what fits their situation best.
Otherwise, the schema seems to be derived from the class being serialized for typed languages, or otherwise annotated in code. The serializer and deserializer code must be manually written to be compatible instead of both sides being codegen'd to match from a schema file. He's the example I found for python: https://fory.apache.org/docs/docs/guide/python_serialization...
You don’t need to hand‑write serializer code. In typed languages you just define your class or struct as usual; in dynamic languages you can use type hints.
When running in compatible mode, Fory automatically derives a compact schema from those definitions at runtime time and sends it along to peers for the first time serialization. That way, both sides know the structure without needing a separate schema file.
The idea is to make cross‑language exchange work out‑of‑the‑box, while still allowing teams to add an explicit IDL later if they want a single source of truth.
Exactly, i'm using protobuf, i've got those proto files which generate code in python, ruby, go and typescript. I can share the package into those different services and i'm gonna be sure i'll use common interface.
It's not clear for me how to achieve the same with fory?
I get your concern — for one or two languages, skipping an IDL can work well and keeps things simple.
But once you’re dealing with three or more languages, I agree an IDL becomes valuable as a single source of truth. That’s work we’ve started: adding optional IDL support so teams can generate data structures in each language from one shared definition.
Why this over serialization free formats like CapnProto and Flatbuffers? If you want it to be compact, send it through zstd (with a custom dictionary).
I do really like that is broad support out of the box and looks easy to use.
For Python I still prefer using dill since it handles code objects.
I also did a benchmark with dill, it shows that: fory is 20~40X faster and up to 7x higher compression ratio compared to dill. I don't dive into dill to see how it works. Here is my benchmark code:
Still mad they had to change the name. "Fury" was a really fitting name for fast serialization framework, "fory" is just bogus. Should've renamed it to "foray" or something.
These binary protocols generally also try to keep the data size small. Protobuf is essentially compressing its integers (varint or zigzag encoding), for example.
It'd be helpful to see a plot of serialization costs vs data size. If you only display serialization TPS, you're always going to lose to the "do nothing" option of just writing your C structs directly to the wire, which is essentially zero cost.
Probably not for everyone. The current limit of 4096 types could be expanded if there’s a real need — it’s not a hard technical barrier.
I’m curious though: what’s an example scenario you’ve seen that requires so many distinct types? I haven’t personally come across a case with 4,096+ protocol messages defined.
If your binary has a small function set, probably not. But in a use case if you want to proxy/intercept cloud APIs, then something like Google APIs has 34K message types:
I think this more speaks to the tradeoff of not having an IDL where the deserializer either knows what type to expect if it was built with the IDL file version that defined it, e.g., this recent issue:
If schema consistent mode is enabled globally when creating fory, type meta will be written as a fory unsigned varint of type_id. Schema evolution related meta will be ignored.
Curious about comparisons with Apache Arrow, which uses flatbuffers to avoid memory copying during deserialization, which is well supported by the Pandas ecosystem, and which allows users to serialize arrays as lists of numbers that have hardware support from a GPU (int8-64, float)
Apache Arrow is more of a memory format than a general‑purpose data serialization system. It’s great for in‑memory analytics and GPU‑friendly columnar storage.
Apache Fory, on the other hand, has its own wire‑stream format designed for sending data across processes or networks. Most of the code is focused on efficiently converting in‑memory objects into that stream format (and back) — with features like cross‑language support, circular reference handling, and schema evolution.
Fory also has a row format, which is a memory format, and can complement or compete with Arrow’s columnar format depending on the use case.
We haven’t tested FlatBuffers vs. Fory for Rust yet, but we plan to.
It’s also worth noting the focus is a bit different: FlatBuffers is great for certain constrained scenarios (e.g., games, embedded), while Fory is a more general‑purpose serialization framework with features like cross‑language object graph support, circular reference handling, and schema evolution.
No, it's not needed. We plan to remove Google Guava from the Fory Java dependency. Our philosophy is that the core should have as few dependencies as possible for maintainability and minimal footprint.
What's the story for JS. I see that there is a javascript directory, but it only mentions nodejs. I don't see an npm package. So does this work in web browsers?
I understand your concern. Personally, I’m open to using AI as part of the development process, but the majority of the code in this repository has been written manually and carefully reviewed.
The AGENTS.md file was added only recently. Its original purpose wasn’t to hand code generation entirely to an AI, but rather to use AI as a pair‑programming debugger — for example, having it walk through tricky binary parsing issues byte‑by‑byte. Serialization in a compact binary format can be hard to debug, and AI can sometimes save hours by quickly pinpointing structural mismatches.
That said, serialization is a fundamental piece of infrastructure, so we remain conservative: any AI‑assisted changes go through the same rigorous review and test process as everything else. As technology evolves, I think it’s worth exploring new tools — but with care.
The https://github.com/apache/fory/blob/main/AGENTS.md is a very detailed document only for AI coding, but an excelent reference for development. But you are right, it may introduce concerns, let me remove it from landing pages
I wish we would focus on making tooling better for W3C EXI (Binary XML encoding) instead of inventing new formats. Just being fast isn't enough, I don't see many using Aeron/SBT, it need a ecosystem - which XML does have.
Binary XML encoding (like W3C EXI) is useful in some contexts, but it’s generally not as efficient as modern binary serialization formats. It also can’t naturally express shared or circular reference semantics, which are important for complex object graphs.
Fory’s format was designed from the ground up to handle those cases efficiently, while still enabling cross‑language compatibility and schema evolution.
I am not sure if W3C EXI, or ASN.1 BER or something else is better, but agree that using DOP (rather than OOP) design principles is the right answer -- which means focusing on the encoding first, and working backwards towards the languages / clients.
DOP is great, but there’s always a gap between DOP and OOP. That gap is where Fory comes in. Right now, Fory takes an OOP‑first approach, but next we’ll add a DOP path by introducing an optional IDL — bridging the two styles. My goal is for the IDL to also support optional OOP‑style expressiveness, so teams can choose the balance that fits their needs.
DOP is very interesting, I like this idea too — most DOP approaches are implemented via an IDL, which is another valid direction. I plan to support that in Fory. I want to give users the freedom to choose the model that works best for them.
Are the benchmarks actually fair? See:
https://github.com/apache/fory/blob/fd1d53bd0fbbc5e0ce6d53ef...
It seems if the serialization object is not a "Fory" struct, then it is forced to go through to/from conversion as part of the measured serialization work:
https://github.com/apache/fory/blob/fd1d53bd0fbbc5e0ce6d53ef...
The to/from type of work includes cloning Strings:
https://github.com/apache/fory/blob/fd1d53bd0fbbc5e0ce6d53ef...
reallocating growing arrays with collect:
https://github.com/apache/fory/blob/fd1d53bd0fbbc5e0ce6d53ef...
I'd think that the to/from Fory types is shouldn't be part of the tests.
Also, when used in an actual system tonic would be providing a 8KB buffer to write into, not just a Vec::default() that may need to be resized multiple times:
https://github.com/hyperium/tonic/blob/147c94cd661c0015af2e5...
IMO, not a fair benchmark.
I can see the source of an 10x improvement on an Intel(R) Xeon(R) Gold 6136 CPU @ 3.00GHz, but it drops to 3x improvement when I remove the to/from that clones or collects Vecs, and always allocate an 8K Vec instead of a ::Default for the writable buffer.
If anything, the benches should be updated in a tower service / codec generics style where other formats like protobuf do not use any Fory-related code at all.
Note also that Fory has some writer pool that is utilized during the tests:
https://github.com/apache/fory/blob/fd1d53bd0fbbc5e0ce6d53ef...
Original bench selection for Fory:
Compared to original bench for Protobuf/Prost: However after allocating 8K instead of ::Default and removing to/from it for an updated protobuf bench:The Rust benchmarks in Fory are intended more as end‑to‑end benchmarks for typical OOP‑style application scenarios, not just raw buffer write speed.
Protobuf is very much a DOP (data‑oriented programming) approach — which is great for some systems. But in many complex applications, especially those using polymorphism, teams don’t want to couple Protobuf‑generated message structs directly into their domain models. Generated types are harder to extend, and if you embed them everywhere (fields, parameters, return types), switching to another serialization framework later becomes almost impossible without touching huge parts of the codebase.
In large systems, it’s common to define independent domain model structs used throughout the codebase, and only convert to/from the Protobuf messages at the serialization boundary. That conversion step is exactly what’s represented in our benchmarks — because it’s what happens in many real deployments.
There’s also the type‑system gap: for example, if your Rust struct has a Box<dyn Trait> field, representing that cleanly in Protobuf is tricky. You might fall back to a oneof, but that essentially generates an enum variant, which often isn’t what users actually want for polymorphic behavior.
So, yes — we include the conversion in our measurements intentionally, to reflect the real‑world large systems practices.
Yes, I agree that protos usually should only be used at the serialization boundary, as well as the slightly off-topic idea that the generated code should be private to the package and/or binary.
So to reflect the real‑world practices, the benchmark code should then allocate and give the protobuf serializer an 8K Vec like in tonic, and not an empty one that may require multiple re-allocations?
Yes, you are right. I missed this when I wrote the benchmark code. I will update the benchmark code and share the results here
DOP is great for certain scenarios, but there’s always a gap between DOP and OOP. That gap is where an extra domain model and the conversion step come in — especially in systems that rely heavily on polymorphism or want to keep serialization types decoupled from core business models.
Regarding design tradeoffs: I am very skeptical that this can be made to work for the long run in a cross-language way without formalizing the on-the-wire contract via IDL or similar.
In my experience, while starting from a language to arrive at the serialization often feels more ergonomic (e.g. RPC style) in the start, it hides too much of what's going on from the users and over time suffers greatly from programming language / runtime changes - the latter multiplied by the number of languages or frameworks supported.
That’s a fair point — with more languages in the mix, having a formal schema can definitely help prevent drift.
The way I think about it is: • Single‑language projects often work best without an IDL — it keeps things simple and avoids extra steps. • Two languages – both IDL and no‑IDL approaches can work, depending on the team’s habits. • Three or more – an IDL can be really useful as a single source of truth and to avoid manually writing struct definitions in every language.
For Apache Fory, my plan is to add optional IDL support, so teams who want that “single truth” can generate definitions automatically, and others can continue with language‑first development. My hope is to give people flexibility to choose what fits their situation best.
even in single language, how do you handle schema evolution (detect breaking changes at compile time) without IDL?
I'm wondering how do you share you shared types between languages if there's no schema ?
Looks like there's a type mapping chart for supported types: https://fory.apache.org/docs/docs/guide/xlang_type_mapping
Otherwise, the schema seems to be derived from the class being serialized for typed languages, or otherwise annotated in code. The serializer and deserializer code must be manually written to be compatible instead of both sides being codegen'd to match from a schema file. He's the example I found for python: https://fory.apache.org/docs/docs/guide/python_serialization...
You don’t need to hand‑write serializer code. In typed languages you just define your class or struct as usual; in dynamic languages you can use type hints.
When running in compatible mode, Fory automatically derives a compact schema from those definitions at runtime time and sends it along to peers for the first time serialization. That way, both sides know the structure without needing a separate schema file.
The idea is to make cross‑language exchange work out‑of‑the‑box, while still allowing teams to add an explicit IDL later if they want a single source of truth.
He's saying you need to hand-write the class or struct. You don't need to do that in systems with an explicit schema like Protobuf.
Exactly, i'm using protobuf, i've got those proto files which generate code in python, ruby, go and typescript. I can share the package into those different services and i'm gonna be sure i'll use common interface.
It's not clear for me how to achieve the same with fory?
I am confused on this as well, they list polyglot teams[0] as their top use case and consider not needing schema files a feature
[0] https://fory.apache.org/blog/2025/10/29/fory_rust_versatile_...
I am skeptical that it's possible to make this work in the long run.
I get your concern — for one or two languages, skipping an IDL can work well and keeps things simple.
But once you’re dealing with three or more languages, I agree an IDL becomes valuable as a single source of truth. That’s work we’ve started: adding optional IDL support so teams can generate data structures in each language from one shared definition.
Not explaining this case makes me wonder how much this lib is actually used in production. This was also the first question I asked myself.
This is our first release for Fory Rust. The Java and Python bindings has been used widely. You can see from https://fory.apache.org/user
Why this over serialization free formats like CapnProto and Flatbuffers? If you want it to be compact, send it through zstd (with a custom dictionary).
I do really like that is broad support out of the box and looks easy to use.
For Python I still prefer using dill since it handles code objects.
https://github.com/uqfoundation/dill
I also did a benchmark with dill, it shows that: fory is 20~40X faster and up to 7x higher compression ratio compared to dill. I don't dive into dill to see how it works. Here is my benchmark code:
https://github.com/chaokunyang/python_benchmarks
Apache Fory is also a drop-in replacement for pickle/cloudpickle, you can use it to serialize code object such as local function/Classes too.
https://github.com/apache/fory/tree/main/python#serialize-lo...
When serializing code objects, pyfory is 3× higher compression ratio compared cloudpickle
And pyfory also provide extra security audit capability to avoid maliciously deserialization data attack.
Benchmark link gives me 404, but I found this link that seems to show the proper benchmarks:
https://fory.apache.org/docs/docs/introduction/benchmark
Still mad they had to change the name. "Fury" was a really fitting name for fast serialization framework, "fory" is just bogus. Should've renamed it to "foray" or something.
I liked the name “Fury” too — I actually named it myself and was really fond of it, but unfortunately we had to change it.
These binary protocols generally also try to keep the data size small. Protobuf is essentially compressing its integers (varint or zigzag encoding), for example.
It'd be helpful to see a plot of serialization costs vs data size. If you only display serialization TPS, you're always going to lose to the "do nothing" option of just writing your C structs directly to the wire, which is essentially zero cost.
Fory also compress integers using varint or zigzag encoding. The size are basically same:
| data type | data size | fory | protobuf |
| --------------- | --------- | ------- | -------- |
| simple-struct | small | 21 | 19 |
| simple-struct | medium | 70 | 66 |
| simple-struct | large | 220 | 216 |
| simple-list | small | 36 | 16 |
| simple-list | medium | 802 | 543 |
| simple-list | large | 14512 | 12876 |
| simple-map | small | 33 | 36 |
| simple-map | medium | 795 | 1182 |
| simple-map | large | 17893 | 21746 |
| person | small | 122 | 118 |
| person | medium | 873 | 948 |
| person | large | 7531 | 7865 |
| company | small | 191 | 182 |
| company | medium | 9118 | 9950 |
| company | large | 748105 | 782485 |
| e-commerce-data | small | 750 | 737 |
| e-commerce-data | medium | 53275 | 58025 |
| e-commerce-data | large | 1079358 | 1166878 |
| system-data | small | 311 | 315 |
| system-data | medium | 24301 | 26161 |
| system-data | large | 450031 | 479988 |
It appears there are two schema compatibility modes and no guarantee of minor version binary compatibility.
Is 4096 types enough for everyone?
https://github.com/apache/fory/blob/fd1d53bd0fbbc5e0ce6d53ef...
Probably not for everyone. The current limit of 4096 types could be expanded if there’s a real need — it’s not a hard technical barrier.
I’m curious though: what’s an example scenario you’ve seen that requires so many distinct types? I haven’t personally come across a case with 4,096+ protocol messages defined.
If your binary has a small function set, probably not. But in a use case if you want to proxy/intercept cloud APIs, then something like Google APIs has 34K message types:
I think this more speaks to the tradeoff of not having an IDL where the deserializer either knows what type to expect if it was built with the IDL file version that defined it, e.g., this recent issue:https://github.com/apache/fory/issues/2818
But now I do see that the 4096 is just arbitrary:
I get a 404 on https://fory.apache.org/docs/benchmarks/rust
You can browse https://fory.apache.org/docs/, but I didn't find any benchmarks directory
Guessing one of these
https://fory.apache.org/docs/docs/introduction/benchmark
https://fory.apache.org/docs/docs/guide/rust_serialization
Curious about comparisons with Apache Arrow, which uses flatbuffers to avoid memory copying during deserialization, which is well supported by the Pandas ecosystem, and which allows users to serialize arrays as lists of numbers that have hardware support from a GPU (int8-64, float)
Apache Arrow is more of a memory format than a general‑purpose data serialization system. It’s great for in‑memory analytics and GPU‑friendly columnar storage.
Apache Fory, on the other hand, has its own wire‑stream format designed for sending data across processes or networks. Most of the code is focused on efficiently converting in‑memory objects into that stream format (and back) — with features like cross‑language support, circular reference handling, and schema evolution.
Fory also has a row format, which is a memory format, and can complement or compete with Arrow’s columnar format depending on the use case.
Would love to see how it compares to Flatbuffers - was surprised to not see it in the benchmarks!
Maybe I'm missing it, but they mention Flatbuffers a lot here, then don't show benchmarks:
https://fory.apache.org/blog/fury_blazing_fast_multiple_lang...
But flatbuffers is _much_ faster than protobuf/json:
https://flatbuffers.dev/benchmarks/
In our Java benchmarks, Fory is actually faster than FlatBuffers — you can try it yourself here: https://github.com/apache/fory/blob/main/java/benchmark/src/...
We haven’t tested FlatBuffers vs. Fory for Rust yet, but we plan to.
It’s also worth noting the focus is a bit different: FlatBuffers is great for certain constrained scenarios (e.g., games, embedded), while Fory is a more general‑purpose serialization framework with features like cross‑language object graph support, circular reference handling, and schema evolution.
> endian flag: 1 when data is encoded by little endian, 0 for big endian.
Have we learned nothing? Endian swap on platforms that need it is faster than conditionals, and simpler.
Is Google guava really needed? I would like it to be taken out.
No, it's not needed. We plan to remove Google Guava from the Fory Java dependency. Our philosophy is that the core should have as few dependencies as possible for maintainability and minimal footprint.
What's the story for JS. I see that there is a javascript directory, but it only mentions nodejs. I don't see an npm package. So does this work in web browsers?
JS support is still experimental, I have not publish it to npm
Looks promising but the level of AI in the repo is a real turn-off.
I understand your concern. Personally, I’m open to using AI as part of the development process, but the majority of the code in this repository has been written manually and carefully reviewed.
The AGENTS.md file was added only recently. Its original purpose wasn’t to hand code generation entirely to an AI, but rather to use AI as a pair‑programming debugger — for example, having it walk through tricky binary parsing issues byte‑by‑byte. Serialization in a compact binary format can be hard to debug, and AI can sometimes save hours by quickly pinpointing structural mismatches.
That said, serialization is a fundamental piece of infrastructure, so we remain conservative: any AI‑assisted changes go through the same rigorous review and test process as everything else. As technology evolves, I think it’s worth exploring new tools — but with care.
You mean this file? https://github.com/apache/fory/blob/main/AGENTS.md
Yes, but also the README is clearly heavily AI derived.
How does this deal with numeric types like NaN, Infinity...?
link is 404 for me
should be ok now
The prevalence of AI slop in the landing page doc does not inspire confidence.
The https://github.com/apache/fory/blob/main/AGENTS.md is a very detailed document only for AI coding, but an excelent reference for development. But you are right, it may introduce concerns, let me remove it from landing pages
That’s not what I meant, I mean it is obvious from many phrases on the landing page that they were written with AI