My concern is, if decode fails I need to debug WASM added by some other party maybe containing random bugs. Maybe a library of standard decoders maintained and tested by the project could help, but then not sure if it kills the advantage of the flexibility it provides.
This bit is quite genius, rather than depend on a language-specific SDK/lib for working with the formats you can fallback to exported WASM methods if none exist:
> "Each self-describing F3 file includes both the data and meta-data, as well as WebAssembly (Wasm) binaries to decode the data. Embedding the decoders in each file requires minimal storage (kilobytes) and ensures compatibility on any platform in case native decoders are unavailable. "
WASM has strong tried and proven sandboxing. We basically can build on nearly 30 years of experience. The decoders don't need a lot of access, they can basically be pure functions.
If this will pan out security-wise I don't know. I'm more worried that it will be so slow that no one will use it. Interesting idea, though, and I can see applications outside of the "big data" realm this apparently targets.
> The decoders don't need a lot of access, they can basically be pure functions
They don't currently either do they? It's the tight coupling of the interface layer no? I'm not sure this would be faster, or more secure so reliability might be the best usecase?
It sounds neat, but feels like it might fall apart with higher-complexity formats. What does an embedded decoder for a PDF look like? I guess since they are tightly-coupled to the file bytes themselves, the author of the file gets to choose what formats make sense, but not all formats have a one-true-decode-step.
Despite the name seemingly implying otherwise, F3 is an alternative to columnar storage formats like Parquet; the goal is not to support every conceivable encoding of every file type such as a PDF. Think of the use cases being more like "What if you used a specialized compressor and need a custom block decompression algorithm" or "Decode internal format into Arrow output" or something like that.
That would be why it chose a VM that is explicitly designed for sandboxing rather than native executable code or similar, the risk can be minimized by reducing the surface area available to that executable code to almost nothing.
TrueType and OpenType fonts include code executed by a VM to even render them. This wasn't a viable source of attacks so far, due to the properly limited nature of the VMs.
Maybe I would pick the eBPF VM instead, with all its limiting and verifying mechanics.
There are many documented, exploited-in-the-wild font-file attacks (one example in 1]). Apple is re-writing their font interpreter specifically to improve security. [2]
> This security update resolves a publicly disclosed vulnerability in Microsoft Windows. The vulnerability could allow remote code execution if a user opens a specially crafted document or visits a malicious Web page that embeds TrueType font files.
> This security update is rated Critical for all supported releases of Microsoft Windows. For more information, see the subsection, Affected and Non-Affected Software, in this section.
> The security update addresses the vulnerability by modifying the way that a Windows kernel-mode driver handles TrueType font files. For more information about the vulnerability, see the Frequently Asked Questions (FAQ) subsection for the specific vulnerability entry under the next section, Vulnerability Information.
Sure, but when the standard says "read this file and execute the instructions you find at the beginning" that is more dangerous than "this is a file with data and your program needs to figure out how to read it".
Ok if you want to be pedantic, the standard says, "if you can't read this file, go to the offset and then execute the code you find" which isn't functionally different from what I said.
There’s a big difference in the expected use of a file. If the file is attacker provided, and the fallback path is being used, the attacker can embed whatever WASM payload they want into the file since the file will be “opened” by “execute this offset into the file”.
Compare that to JSON. The parser NEVER needs to execute arbitrary instructions. Parser might have bugs, but it avoids a whole class of issues.
Putting aside the WASM sandboxing (I’m not familiar enough with it to understand how sandboxing works) there’s a DoS vector at least. Even regexes have had many DoS issues, and I can’t imagine WASM being easier to sandbox for DoS risk.
gotcha, so the vulnerability will be in some common libraries that attackers force some wasm fallback path with custom wasm instructions that when executed does something nefarious.
.exe has bindings to OS ABI and system calls, WASM doesn’t have this by default, it’s up to the VM to provide whatever environment the WASM executable needs, ideally there should be no system calls, no stdio, just instructions on how to interpret the file format.
Double-clicking an ".exe" (or running it via a shell) is not the same as "bag of bytes", it's "send these bytes to an executable environment".
Doing `head foo.exe` is quite different than `run foo.exe`
If I encode executable instructions in "image.png" and then send them to an interpreter that runs those instructions, the file extension doesn't matter.
I would call it clever. I'm not sure I'd call it genius.
When I'm working with data I'm working in a specific set of languages. Usually one. Yeah, other people might be working in other languages, but no individual author really needs a language-agnostic way of accessing data beyond compile time. Add to that the likely runtime boundaries that may need to be crossed instead of e.g. inlined by the compiler because it's in-language and dealing with known offsets or tags (depends on the data format of course). To the other commenter's point, am I going to have to sandbox all data access code just to be sure it's not able to do something unexpected? There's a lot of complexity here. And the inherent risk is going to slow down the operation that should be the simplest and fastest: interpreting bytes.
A big problem with parquet, which this aims to replace, is that it's hard to add new encodings because everyone wants to stay compatible with old readers. Embedding the decoders in the file as WASM solves this problem since in theory, old readers will be able to read new files by just using the provided WASM to decode a column whose format the reader doesn't recognize.
So this is really about making a file that is forwards compatible in a way that lets you push the standards more than existing formats.
Hm. I can kinda see it replacing self-extracting EXEs, but a lot of why you choose specific file formats is for specific features they offer - any self-describing system can fall into "there are too many competing features and nobody handles them all" exactly as easily as any other format.
Like, can this file be efficiently mmap'd? Maybe if it emulates tar internally, but you don't know until you run it. Can it be seeked to specific bytes too only decompress part? It only supports a pre-release version of ISO-36898533 seeking, and your file library dropped support for it 6 years ago. If I rewrite 1MB in the middle, can it only change those pages on disk (and maybe an index), or do I have to rewrite the whole thing? Well the wasm blob supports 97 different APIs for it, so it's larger than the data (but nobody paid attention to that), so you have 19 options that you recognize, but your CPU's native WASM only accelerates two or three so you've still got to specialize your code heavily.
At least with "*.tar" you have some idea of what's possible.
One nice thing about some modern formats is that there are tools that read them at extraordinarily high effective speed. For example, DuckDB can do all manner of nifty optimizations while reading its own native format or Parquet. And I’m not sure that those optimizations can be effectively applied to a format that needs a WASM blob to be run to understand it. By the time you run a non-SIMD or even a SIMD-optimized pass over app the data, if that pass doesn’t understand your query, you may have already lost.
I admit I only skimmed the beginning of the paper, and maybe the format is less general than it sounds.
Some folks described it as genius. My initial reaction that it's somewhat silly. A major aspect of any data format is what you do with the data once decoded. An audio file represents something completely different than a HTML document, which is completely different from model weights. An embedded VM that convers one bit stream to another solves just a small piece of the puzzle. It doesn't let you play videos with a text editor. So what's the goal?
Backward compatibility doesn't seem like a worthwhile use case, so I guess the main one is forward compatibility: say I come up with a video compression scheme that's better than H.265, but not all platforms support it, so I embed a decoder that would allow me to play it back on legacy hardware. But I guess that also shows the weakness of the idea: it's unlikely that legacy hardware will perform well doing software-only decode for video formats from the future.
It's also a maintenance nightmare: if your decoder has a bug that needs fixing, how do you patch all the files that already embed it?
All in all, it sounds like we're adding megabytes of code and a considerable attack surface to every format parser. It's more opportunities for remote code execution, resource exhaustion attacks, and so on. What do we realistically get in return?
Nice! The world can always use a better data format.
I think you might get some traction if you post the advantages over parquet and other files directly on the readme, so that if someone goes to https://github.com/future-file-format/f3 the see why they should try it.
Mention the advantages and post metrics. Cherry pick the metrics! There's probably a good use case for this but, from the current readme, it's not clear who should use this and why.
If I am archiving PBs of data for 10+ years, I don't want to rely on a WASM interpreter being available and performant in the future just to read a file. I want a dead-simple, heavily documented byte specification like Parquet.
Additionally, putting the decoding logic inside an WASM binary introduces an active execution layer into what should be a cold storage.
The description mentions shortcomings of the previous file types like parquet, but it isn't really evident to me what those shortcomings are, or if the use cases for parquet and F3 have really that much of an overlap to make this comparison valid in the first place.
To save a click it's a file format for columnar data specifically (like Parquet), which they very generically named Future-proof File Format. Most of this could fit in the title instead of just "F3" that is a bit of bait.
It doesn't explain what the project does (a file format for what? Name dropping other things I haven't heard of isn't useful)
There are no examples. It links to a flatbuffer schema which is at least well commented, but is full of deep implementation details.
The point is that within 2-3 minutes I'm not convinced why I care and still don't know enough about what this is to even think back to if if I encounter a scenario in the future where it would be useful.
> designed with efficiency, interoperability, and extensibility in mind. It provides a data organization that rectifies the layout shortcomings of the last-generation formats like Parquet,
This is all marketing speak that says nothing.
> maintaining good interoperability and extensibility (a.k.a future-proof) via embedded Wasm decoders
What does this even mean? Providing a decoder is no guarantee of futureproofness.
I love the idea, and I developed something similar of myself in the past (https://github.com/golddranks/kobuta), but... this reeks of slop. With Rust code, edition="2021" is a dead giveaway.
My concern is, if decode fails I need to debug WASM added by some other party maybe containing random bugs. Maybe a library of standard decoders maintained and tested by the project could help, but then not sure if it kills the advantage of the flexibility it provides.
This bit is quite genius, rather than depend on a language-specific SDK/lib for working with the formats you can fallback to exported WASM methods if none exist:
So attackers don't have to craft specially corrupted files? They can just include the code to perform the attack in the data file itself?
WASM has strong tried and proven sandboxing. We basically can build on nearly 30 years of experience. The decoders don't need a lot of access, they can basically be pure functions.
If this will pan out security-wise I don't know. I'm more worried that it will be so slow that no one will use it. Interesting idea, though, and I can see applications outside of the "big data" realm this apparently targets.
How do you prevent compression bomb attacks when files can define their own compression functions?
You could have some kind of OOM killer, but that will be a "footgun" that people who are actually doing "big data" will constantly shoot.
This pretty much kills any ingestion pipeline where the source is untrusted.
> The decoders don't need a lot of access, they can basically be pure functions
They don't currently either do they? It's the tight coupling of the interface layer no? I'm not sure this would be faster, or more secure so reliability might be the best usecase?
Yes...my first thought. No way in hell anyone actually trusts this.
(And as if we didn't trust the compiler enough already!)
Does WASM have built-in I/O? If not, all that a decoder would be able to do is to decode into a buffer.
But the WASM runs in the sandbox! It only has access to some files, your display, inputs, ... nothing insecure at all!
WASM runs in a confined memory space allocated for the program. There is no I/O or host address space access.
You need to run a WASI environment for that.
It sounds neat, but feels like it might fall apart with higher-complexity formats. What does an embedded decoder for a PDF look like? I guess since they are tightly-coupled to the file bytes themselves, the author of the file gets to choose what formats make sense, but not all formats have a one-true-decode-step.
Despite the name seemingly implying otherwise, F3 is an alternative to columnar storage formats like Parquet; the goal is not to support every conceivable encoding of every file type such as a PDF. Think of the use cases being more like "What if you used a specialized compressor and need a custom block decompression algorithm" or "Decode internal format into Arrow output" or something like that.
Applets redux.
How is wasm better than C bindings?
Many languages don't have ergonomic experiences for working with C ABI's without explicit wrapper code.
Hell, Node.js didn't even get this ability until LAST MONTH:
https://nodejs.org/en/blog/release/v26.1.0
You'd have to write a second library to interface the C ABI with Node via NAPI just to consume it.
WASM is platform independent.
What do you mean by C bindings? C bindings to what?
Is embedding executable code into a file a security risk? My assumption is a yes
That would be why it chose a VM that is explicitly designed for sandboxing rather than native executable code or similar, the risk can be minimized by reducing the surface area available to that executable code to almost nothing.
> Is embedding executable code into a file a security risk?
Yes, which is why nobody uses PDFs.
TrueType and OpenType fonts include code executed by a VM to even render them. This wasn't a viable source of attacks so far, due to the properly limited nature of the VMs.
Maybe I would pick the eBPF VM instead, with all its limiting and verifying mechanics.
There are many documented, exploited-in-the-wild font-file attacks (one example in 1]). Apple is re-writing their font interpreter specifically to improve security. [2]
[1] https://www.bleepingcomputer.com/news/security/facebook-disc...
[2] https://blakecrosley.com/blog/truetype-hinting-swift-migrati...
https://learn.microsoft.com/en-us/security-updates/SecurityB...
> This security update resolves a publicly disclosed vulnerability in Microsoft Windows. The vulnerability could allow remote code execution if a user opens a specially crafted document or visits a malicious Web page that embeds TrueType font files.
> This security update is rated Critical for all supported releases of Microsoft Windows. For more information, see the subsection, Affected and Non-Affected Software, in this section.
> The security update addresses the vulnerability by modifying the way that a Windows kernel-mode driver handles TrueType font files. For more information about the vulnerability, see the Frequently Asked Questions (FAQ) subsection for the specific vulnerability entry under the next section, Vulnerability Information.
There is no concept of "executable" vs "non-executable" content in a file.
A file is a bag of bytes. You can send those bytes to different things, like a text editor's content-stream, or as the input to a WASM interpreter.
What you decide to do with the bytes in a file is your own prerogative. Each byte is whatever you make of it.
Sure, but when the standard says "read this file and execute the instructions you find at the beginning" that is more dangerous than "this is a file with data and your program needs to figure out how to read it".
I guess it's a good thing that the F3 standard does not say "read this file and execute the instructions you find at the beginning", then?
The WASM encoders/decoders are embedded resources that exist as byte offsets in the file metadata, not header info.
Ok if you want to be pedantic, the standard says, "if you can't read this file, go to the offset and then execute the code you find" which isn't functionally different from what I said.
There’s a big difference in the expected use of a file. If the file is attacker provided, and the fallback path is being used, the attacker can embed whatever WASM payload they want into the file since the file will be “opened” by “execute this offset into the file”.
Compare that to JSON. The parser NEVER needs to execute arbitrary instructions. Parser might have bugs, but it avoids a whole class of issues.
WASM physically cannot interact with the underlying host or perform I/O -- you need a WASI environment for that.
Putting aside the WASM sandboxing (I’m not familiar enough with it to understand how sandboxing works) there’s a DoS vector at least. Even regexes have had many DoS issues, and I can’t imagine WASM being easier to sandbox for DoS risk.
There exist Wasm interpreters capable of limiting the number of instructions executed.
gotcha, so the vulnerability will be in some common libraries that attackers force some wasm fallback path with custom wasm instructions that when executed does something nefarious.
I'd say at worst it's setup for poor security
I mean can't we say the same thing about sending around a .exe though?
.exe has bindings to OS ABI and system calls, WASM doesn’t have this by default, it’s up to the VM to provide whatever environment the WASM executable needs, ideally there should be no system calls, no stdio, just instructions on how to interpret the file format.
Double-clicking an ".exe" (or running it via a shell) is not the same as "bag of bytes", it's "send these bytes to an executable environment".
Doing `head foo.exe` is quite different than `run foo.exe`
If I encode executable instructions in "image.png" and then send them to an interpreter that runs those instructions, the file extension doesn't matter.
exactly
I would call it clever. I'm not sure I'd call it genius.
When I'm working with data I'm working in a specific set of languages. Usually one. Yeah, other people might be working in other languages, but no individual author really needs a language-agnostic way of accessing data beyond compile time. Add to that the likely runtime boundaries that may need to be crossed instead of e.g. inlined by the compiler because it's in-language and dealing with known offsets or tags (depends on the data format of course). To the other commenter's point, am I going to have to sandbox all data access code just to be sure it's not able to do something unexpected? There's a lot of complexity here. And the inherent risk is going to slow down the operation that should be the simplest and fastest: interpreting bytes.
A big problem with parquet, which this aims to replace, is that it's hard to add new encodings because everyone wants to stay compatible with old readers. Embedding the decoders in the file as WASM solves this problem since in theory, old readers will be able to read new files by just using the provided WASM to decode a column whose format the reader doesn't recognize.
So this is really about making a file that is forwards compatible in a way that lets you push the standards more than existing formats.
Hm. I can kinda see it replacing self-extracting EXEs, but a lot of why you choose specific file formats is for specific features they offer - any self-describing system can fall into "there are too many competing features and nobody handles them all" exactly as easily as any other format.
Like, can this file be efficiently mmap'd? Maybe if it emulates tar internally, but you don't know until you run it. Can it be seeked to specific bytes too only decompress part? It only supports a pre-release version of ISO-36898533 seeking, and your file library dropped support for it 6 years ago. If I rewrite 1MB in the middle, can it only change those pages on disk (and maybe an index), or do I have to rewrite the whole thing? Well the wasm blob supports 97 different APIs for it, so it's larger than the data (but nobody paid attention to that), so you have 19 options that you recognize, but your CPU's native WASM only accelerates two or three so you've still got to specialize your code heavily.
At least with "*.tar" you have some idea of what's possible.
One nice thing about some modern formats is that there are tools that read them at extraordinarily high effective speed. For example, DuckDB can do all manner of nifty optimizations while reading its own native format or Parquet. And I’m not sure that those optimizations can be effectively applied to a format that needs a WASM blob to be run to understand it. By the time you run a non-SIMD or even a SIMD-optimized pass over app the data, if that pass doesn’t understand your query, you may have already lost.
I admit I only skimmed the beginning of the paper, and maybe the format is less general than it sounds.
My understanding is it’s a fallback mechanism
Some folks described it as genius. My initial reaction that it's somewhat silly. A major aspect of any data format is what you do with the data once decoded. An audio file represents something completely different than a HTML document, which is completely different from model weights. An embedded VM that convers one bit stream to another solves just a small piece of the puzzle. It doesn't let you play videos with a text editor. So what's the goal?
Backward compatibility doesn't seem like a worthwhile use case, so I guess the main one is forward compatibility: say I come up with a video compression scheme that's better than H.265, but not all platforms support it, so I embed a decoder that would allow me to play it back on legacy hardware. But I guess that also shows the weakness of the idea: it's unlikely that legacy hardware will perform well doing software-only decode for video formats from the future.
It's also a maintenance nightmare: if your decoder has a bug that needs fixing, how do you patch all the files that already embed it?
All in all, it sounds like we're adding megabytes of code and a considerable attack surface to every format parser. It's more opportunities for remote code execution, resource exhaustion attacks, and so on. What do we realistically get in return?
This could use a bit more "why".
Shortcomings of Parquet are mentioned as overcome by this, which ones? Certainly not wide tool support...
Why should one leave Parquet or ORC for this structure?
The ‘why’ is referenced in the bibliography at the end of the readme; this repo is not meant to be consumed standalone. Start with the paper instead:
https://doi.org/10.1145/3749163
Paper mentions Parquet, ORC, Nimble, Lance, TSFile, Bullion, and BtrBlocks.
Yeah it seems like most of this can be handled by some more dev hours to Parquet
Nice! The world can always use a better data format.
I think you might get some traction if you post the advantages over parquet and other files directly on the readme, so that if someone goes to https://github.com/future-file-format/f3 the see why they should try it.
Mention the advantages and post metrics. Cherry pick the metrics! There's probably a good use case for this but, from the current readme, it's not clear who should use this and why.
If I am archiving PBs of data for 10+ years, I don't want to rely on a WASM interpreter being available and performant in the future just to read a file. I want a dead-simple, heavily documented byte specification like Parquet.
Additionally, putting the decoding logic inside an WASM binary introduces an active execution layer into what should be a cold storage.
You don't want to run a custom 10-year-old data parsing function every time you read a single data record?
File format for what? Text, graphics, compiled code?
For columnar data storage I think. The description references Parquet and they appear to benchmark against Parquet, Vortex, and Lance.
The future.
Great! I'll use it.
In the "future."
Nimble? Lance? Also in the future. Maybe.
I'll use Parquet in the present.
The description mentions shortcomings of the previous file types like parquet, but it isn't really evident to me what those shortcomings are, or if the use cases for parquet and F3 have really that much of an overlap to make this comparison valid in the first place.
To save a click it's a file format for columnar data specifically (like Parquet), which they very generically named Future-proof File Format. Most of this could fit in the title instead of just "F3" that is a bit of bait.
This project README is not particularly useful:
It doesn't explain what the project does (a file format for what? Name dropping other things I haven't heard of isn't useful)
There are no examples. It links to a flatbuffer schema which is at least well commented, but is full of deep implementation details.
The point is that within 2-3 minutes I'm not convinced why I care and still don't know enough about what this is to even think back to if if I encounter a scenario in the future where it would be useful.
> designed with efficiency, interoperability, and extensibility in mind. It provides a data organization that rectifies the layout shortcomings of the last-generation formats like Parquet,
This is all marketing speak that says nothing.
> maintaining good interoperability and extensibility (a.k.a future-proof) via embedded Wasm decoders What does this even mean? Providing a decoder is no guarantee of futureproofness.
Tabular data, it wants to replace Parquet
Probably not a good idea to name your project “future” anything, if you expect that future to become the present.
Also, f3 is already “fight-flash-fraud”.
>via embedded Wasm decoders
runs screaming
I love the idea, and I developed something similar of myself in the past (https://github.com/golddranks/kobuta), but... this reeks of slop. With Rust code, edition="2021" is a dead giveaway.
Don't know why but I had to think of https://xkcd.com/2116/
No commits in 8 months?
Yeah this was a research project, it doesn't look like this is getting any adoption
Is what?
A more descriptive title would be helpful OP:
F3: Open-source data file format for the future
Previous discussion:
2025 https://news.ycombinator.com/item?id=45437759
https://xkcd.com/927/