This is pretty awesome. I work with editors and monaco-like things a ton, and I review (look at) very large PRs very often. Having this speedy optimized interface is a delight. Check out their trees lib as well.
It's cool seeing all the engineering that goes into optimizing performance of diffs. I'm working on a FreeCAD workbench that generates diffs on CAD model trees[0], and although my bottlenecks are a bit different, I can still implement some of your optimizations down the line if needed (such as deferred syntax highlighting).
My main bottleneck is that I do a complete diff on all open + changed documents in the repository up front, because due to how document properties are stored, I won't know if the file has meaningful changes until I compute the full diff (FreeCAD may save the document, but not have anything meaningful change.)
What an interesting article. I did not assume I would read it until the end when I opened it, but the writing was super clear and easy to follow.
At the end, I admire the craft and patience to try to solve code diff rendering, and wish the folks at GitHub could put the same effort to improve their platform.
On a side note, I feel that we’re going to see more and more of this type of agentic usage, in well defined sub tasks, and the ability of a model to try many possibilities is a huge gift here.
I don't understand the point of the inverse sticky technique. Scrolling too fast still breaks the experience (content refuses to scroll), and in a way that, at least to me, feels more disruptive than blanking for a fraction of a second. I might just be too used to blanking.
Also ... shouldn't browsers just be able to render the diff without any of the trickery? Is the browser's job actually that hard for long pages, or are they just not optimising for this? Or is there some other reason for the virtualisation (e.g. memory usage)?
I probably didn’t explain this well enough, but your render times always have to be within the frame buffer (16.6ms for 60hz or 8.3ms for 120hz). Under normal circumstances even if you occasionally blow a frame buffer, with the over-scroll you won’t hit the sticky bounds.
The only time you will is if you’re scrolling at a rate where the jumps are quite large — large and fast enough typically where you’re not going to have a frame of reference for what you should see vs what you are seeing to notice you are behind.
Ultimately scrolling is managed on a separate thread from JS, which means if you do like an opt+click on the scroll bar, you’re going to make a jump that JavaScript can never keep up with, even if you’re under your frame times.
And with regards to safari, if your requestAnimationFrame is capped at 60hz but your scrolling is GPU composited at 120hz, this is the only way to keep scrolling at 120hz with 60hz dom updates and never see any blanking.
hmm it does seem like a good point that browsers should be able to do this natively. the DOM already holds the content, and browsers manage what's rendered in the viewport.
Feature request; `git diff --color-moved` uses colors to display moved chunks of code. Scanning https://diffs.com/docs it isn't obvious that yall support that; please add it :)
It is SO NICE to see people working on making fast, nice-to-use tools. It's a lovely experience to use diffshub. Thank you for creating it, and than you for the great write-up! (I made it "that far" )
I disagree with the theory that scrolling frame rate doesn't need to be smooth for scrolling to feel smooth.
On mobile it kinda does. Scrolling diffs on mobile just kinda feels crap.
I have been spoiled by years of engineer hours spent getting scrolling to be 60- or even 120Hz smooth to match my finger, and diffs just.. isn't.
I know this is frustrating to hear, and that this is technically compounded by mobile probably having the lowest device performance to be playing with too, but.. There you go.
> disagree with the theory that scrolling frame rate doesn't need to be smooth for scrolling to feel smooth
It's possible you might be misunderstanding what I was trying to say here because 120hz scrolling on a 120hz device was the goal and why one of those virtualization techniques was not acceptable to me which lead me to coming up with a novel workaround to this problem (Inverse Sticky Technique).
CodeView uses a system that allows scrolling to update at your native framerate (120hz) WITHOUT needing Javascript needing to keep up at 120hz. If you're seeing stuttering while scrolling on https://diffshub.com would love to know more context (device/diff link/etc) because that is very much NOT our experience.
Even the linked ghostty PR on your home shows this - this is Firefox Android on a Nokia XR21 / TA-1486.
It's not unuseable, but it definitely feels like 'js hacking my scrolling' and not a native surface flinging around.
The experience is actually worse with smaller movements, i guess because my brain is more conscious when breaking the 'finger physically moving the text' illusion.
I don't mean to be dismissive - you're working on a really hard problem, and you're clearly approaching it with a mindset of perfection. I'm posting because I know you're probably able to solve this too :)
Edit: as a point of (unfair) comparison, the codemirror Huge File demo works fine: https://codemirror.net/examples/million/
It does suffer from the occasional partial paint when quickly coasting, but I'm not bothered by this at all, it's far less intrusive than dropping frames / stuttering / etc.
Just to be clear tho, we don't actually scroll jack, native scrolling works as it should and content should move with normal gpu composited scroll. That said, it's possible that loading that much data into memory may be causing causing knock on effects somehow that are just slowing everything down.
Matters a great deal on desktop too, and laptops for that matter. Even more on platforms like macOS that smooths scrolling by default too, but very noticeable on Windows and various Linux distributions too when native scrolling is janky/choppy, and it frustrates even casual users.
I was hoping that this would talk more about the logic behind generating a diff, rather than the optimisations involved in rendering the text.
IMO (as someone who doesn't have to deal with the actual rendering) it would go a bit deeper into talking about deciding how to show what has changed. There's a lot of improvements that could be made there. e.g. "whitespace has changed here" so there's no real code changes involved.
Or "this big list of imports has changed, and code formatting has line-wrapped the list into different lines" - gitlab for example copes poorly with this. I'd love to just see a clean diff that highlights the additional import, and not just ten lines of changes caused by adding one line to a big list of imported symbols/functions.
One of our next big projects is actually to support semantic diffs, which I think will be a lot more applicable to what you're asking for here. Currently diffs just takes a normal git patch file, or generates one from 2 versions of a file.
Most projects start with tree-sitter and then switch to language-native parsers. Either way, it's not something you solve yourself – you just find the language-specific implementation load megabytes of WASM on the frontend or generate it on the backend.
difftastic, semanticdiff.. lots of projects like that. Obviously they can offer stuff like "function name changed" instead of showing you 30 lines of +newName -oldName
> rather than the optimizations involved in rendering the text.
Any views they have on this topic is going to come across as quite opinionated given their choices for text rendering for this post and general aesthetics of website.
Naw, the truth is I'm not really smart or intelligent enough to build a semantic diff system. For that you'll need to wait on a post from one of our smarter devs, this was a post about rendering diffs in a browser.
I remember wrestling with diff tools back in the day. A good visualization can make a world of difference. How does it handle large files with lots of changes?
I feel like virtualization is not the right way to handle things. It adds so much complexity and makes the user experience buggy due to breaking optimizations and features of browsers.
Computers are very powerful these days and have a done of resources that they can use. We should be able to handle large diffs without any crazy tricks.
Keep things as simple as possible and put the whole thing in the DOM. Then if there are performance problems address the scaling problems themselves instead of trying to avoid scaling. For example things like only rendering what is visible should be handled by the browser and not by messing with the DOM. The DOM is not synchronized with the browser's renderer so it will always end up hacky.
You can't say "only rendering what is visible should be handled by the browser" and call that a solution unless you have a magic wand to make Chrome/other browsers do this.
The browser doesn't do this, and so you can either do what you say and have your browser freeze when you load up a million line diff, or you can fix things within your control which is what the author is doing.
Yes and no. It would help to improve things a bit when it comes the measure/reconciliation phase (unclear to say how much). However we've already done a pretty good job around batching writes vs reads.
However passing a million lines of code through pretext is unlikely to be very efficient, so a lot of the work around estimation is still very important.
That said, while I don't want to make pretext a direct dependency of the library, there's a good chance I'll explore the possibility of allowing devs to pass it in as an additional argument perhaps improve performance a bit.
It should also be noted that we have a full API to support things like line annotations (comments, etc) that are entirely controlled by the user, so there's always a bit of a dynamic aspect there that would come into play
Performance and optimization is one of many pieces, but yes, it's a meme to render 500k lines.
That said though, and maybe I didn't say it well in the post, the more performant and optimized your tool is, the less burden you put on developers and users.
Sure you won't review 100k lines, but maybe the diff includes a ton of testing snapshots, or maybe it's a long running feature branch and you need to just quickly jump in and look at a specific change from a specific file. The less the developer or the user needs to think about `how` to render the diff or `how to navigate the diff`, the better we did our job.
Is this really a hard problem? Rendering some text with coloring?
I believe we're just making everything harder than it needs by constraining it to a browser interface. This would be trivial as a terminal application written in native code.
This is pretty awesome. I work with editors and monaco-like things a ton, and I review (look at) very large PRs very often. Having this speedy optimized interface is a delight. Check out their trees lib as well.
For anyone else who's suffering, paste this in the console in devtools:
There's always reader mode when an author tries to be cute about their layout
thank you, how do they live like this.
lmao - thank you!
Aw, Pierre shut down? Is there a write-up on that? (The code review startup idea.)
It's cool seeing all the engineering that goes into optimizing performance of diffs. I'm working on a FreeCAD workbench that generates diffs on CAD model trees[0], and although my bottlenecks are a bit different, I can still implement some of your optimizations down the line if needed (such as deferred syntax highlighting).
My main bottleneck is that I do a complete diff on all open + changed documents in the repository up front, because due to how document properties are stored, I won't know if the file has meaningful changes until I compute the full diff (FreeCAD may save the document, but not have anything meaningful change.)
[0] https://github.com/eblanshey/HistoryWorkbench
What an interesting article. I did not assume I would read it until the end when I opened it, but the writing was super clear and easy to follow.
At the end, I admire the craft and patience to try to solve code diff rendering, and wish the folks at GitHub could put the same effort to improve their platform.
On a side note, I feel that we’re going to see more and more of this type of agentic usage, in well defined sub tasks, and the ability of a model to try many possibilities is a huge gift here.
Hey thank you, appreciate your kind words! I don’t write much and was quite an effort to get all this written out!
I don't understand the point of the inverse sticky technique. Scrolling too fast still breaks the experience (content refuses to scroll), and in a way that, at least to me, feels more disruptive than blanking for a fraction of a second. I might just be too used to blanking.
Also ... shouldn't browsers just be able to render the diff without any of the trickery? Is the browser's job actually that hard for long pages, or are they just not optimising for this? Or is there some other reason for the virtualisation (e.g. memory usage)?
I probably didn’t explain this well enough, but your render times always have to be within the frame buffer (16.6ms for 60hz or 8.3ms for 120hz). Under normal circumstances even if you occasionally blow a frame buffer, with the over-scroll you won’t hit the sticky bounds.
The only time you will is if you’re scrolling at a rate where the jumps are quite large — large and fast enough typically where you’re not going to have a frame of reference for what you should see vs what you are seeing to notice you are behind.
Ultimately scrolling is managed on a separate thread from JS, which means if you do like an opt+click on the scroll bar, you’re going to make a jump that JavaScript can never keep up with, even if you’re under your frame times.
And with regards to safari, if your requestAnimationFrame is capped at 60hz but your scrolling is GPU composited at 120hz, this is the only way to keep scrolling at 120hz with 60hz dom updates and never see any blanking.
hmm it does seem like a good point that browsers should be able to do this natively. the DOM already holds the content, and browsers manage what's rendered in the viewport.
A bit of a technical deep dive into how we built CodeView, a review surface that can handle rendering diffs of immense size, all in a browser.
Feature request; `git diff --color-moved` uses colors to display moved chunks of code. Scanning https://diffs.com/docs it isn't obvious that yall support that; please add it :)
big fan :)
thank you!
It is SO NICE to see people working on making fast, nice-to-use tools. It's a lovely experience to use diffshub. Thank you for creating it, and than you for the great write-up! (I made it "that far" )
Appreciate it, thank you!
I disagree with the theory that scrolling frame rate doesn't need to be smooth for scrolling to feel smooth.
On mobile it kinda does. Scrolling diffs on mobile just kinda feels crap.
I have been spoiled by years of engineer hours spent getting scrolling to be 60- or even 120Hz smooth to match my finger, and diffs just.. isn't.
I know this is frustrating to hear, and that this is technically compounded by mobile probably having the lowest device performance to be playing with too, but.. There you go.
> disagree with the theory that scrolling frame rate doesn't need to be smooth for scrolling to feel smooth
It's possible you might be misunderstanding what I was trying to say here because 120hz scrolling on a 120hz device was the goal and why one of those virtualization techniques was not acceptable to me which lead me to coming up with a novel workaround to this problem (Inverse Sticky Technique).
CodeView uses a system that allows scrolling to update at your native framerate (120hz) WITHOUT needing Javascript needing to keep up at 120hz. If you're seeing stuttering while scrolling on https://diffshub.com would love to know more context (device/diff link/etc) because that is very much NOT our experience.
Even the linked ghostty PR on your home shows this - this is Firefox Android on a Nokia XR21 / TA-1486.
It's not unuseable, but it definitely feels like 'js hacking my scrolling' and not a native surface flinging around.
The experience is actually worse with smaller movements, i guess because my brain is more conscious when breaking the 'finger physically moving the text' illusion.
I don't mean to be dismissive - you're working on a really hard problem, and you're clearly approaching it with a mindset of perfection. I'm posting because I know you're probably able to solve this too :)
Edit: as a point of (unfair) comparison, the codemirror Huge File demo works fine: https://codemirror.net/examples/million/ It does suffer from the occasional partial paint when quickly coasting, but I'm not bothered by this at all, it's far less intrusive than dropping frames / stuttering / etc.
Maybe i need to buy one of these devices to test.
Just to be clear tho, we don't actually scroll jack, native scrolling works as it should and content should move with normal gpu composited scroll. That said, it's possible that loading that much data into memory may be causing causing knock on effects somehow that are just slowing everything down.
Matters a great deal on desktop too, and laptops for that matter. Even more on platforms like macOS that smooths scrolling by default too, but very noticeable on Windows and various Linux distributions too when native scrolling is janky/choppy, and it frustrates even casual users.
I was hoping that this would talk more about the logic behind generating a diff, rather than the optimisations involved in rendering the text.
IMO (as someone who doesn't have to deal with the actual rendering) it would go a bit deeper into talking about deciding how to show what has changed. There's a lot of improvements that could be made there. e.g. "whitespace has changed here" so there's no real code changes involved.
Or "this big list of imports has changed, and code formatting has line-wrapped the list into different lines" - gitlab for example copes poorly with this. I'd love to just see a clean diff that highlights the additional import, and not just ten lines of changes caused by adding one line to a big list of imported symbols/functions.
One of our next big projects is actually to support semantic diffs, which I think will be a lot more applicable to what you're asking for here. Currently diffs just takes a normal git patch file, or generates one from 2 versions of a file.
Most projects start with tree-sitter and then switch to language-native parsers. Either way, it's not something you solve yourself – you just find the language-specific implementation load megabytes of WASM on the frontend or generate it on the backend.
difftastic, semanticdiff.. lots of projects like that. Obviously they can offer stuff like "function name changed" instead of showing you 30 lines of +newName -oldName
> rather than the optimizations involved in rendering the text.
Any views they have on this topic is going to come across as quite opinionated given their choices for text rendering for this post and general aesthetics of website.
Naw, the truth is I'm not really smart or intelligent enough to build a semantic diff system. For that you'll need to wait on a post from one of our smarter devs, this was a post about rendering diffs in a browser.
I remember wrestling with diff tools back in the day. A good visualization can make a world of difference. How does it handle large files with lots of changes?
Semi-related: have you considered making DiffsHub a browser extension, so you can serve private diffs as well?
(I say this, having done a vibe-port of the code to a browser extension, so the underlying concept works.)
When producing TreeTrek, I went with rudimentary diffs that account for colourblind developers:
https://repo.autonoma.ca/repo/treetrek/commit/3fe9360599ae23...
The diffs rendering library looks amazing: https://diffs.com/
Presumably the red-green issue is a simple CSS update?
Yup! You can pick from a bunch of different themes and use css variables to override the core colors as well!
> so hit play on sandstorm
For a brief hopeful moment, I thought this was the .io kind of sandstorm
Very impressive! I doubt Github or Gitlab would ever do something as good as this but maybe there's a chance we could get it in Forgejo?
I feel like virtualization is not the right way to handle things. It adds so much complexity and makes the user experience buggy due to breaking optimizations and features of browsers.
Computers are very powerful these days and have a done of resources that they can use. We should be able to handle large diffs without any crazy tricks.
> I feel like virtualization is not the right way to handle things.
How would you handle it?
Keep things as simple as possible and put the whole thing in the DOM. Then if there are performance problems address the scaling problems themselves instead of trying to avoid scaling. For example things like only rendering what is visible should be handled by the browser and not by messing with the DOM. The DOM is not synchronized with the browser's renderer so it will always end up hacky.
OK but thats simply doesn't perform.
You can't say "only rendering what is visible should be handled by the browser" and call that a solution unless you have a magic wand to make Chrome/other browsers do this.
The browser doesn't do this, and so you can either do what you say and have your browser freeze when you load up a million line diff, or you can fix things within your control which is what the author is doing.
Whatever happened to all the pretext hype? I feel like that would be perfect for rendering huge diffs.
Yes and no. It would help to improve things a bit when it comes the measure/reconciliation phase (unclear to say how much). However we've already done a pretty good job around batching writes vs reads.
However passing a million lines of code through pretext is unlikely to be very efficient, so a lot of the work around estimation is still very important.
That said, while I don't want to make pretext a direct dependency of the library, there's a good chance I'll explore the possibility of allowing devs to pass it in as an additional argument perhaps improve performance a bit.
It should also be noted that we have a full API to support things like line annotations (comments, etc) that are entirely controlled by the user, so there's always a bit of a dynamic aspect there that would come into play
Can you select and copy text with pretext?
Maybe an intended effect, but the header ascii art is suspiciously misaligned... :^)
rendering massive diffs is cool but ultimately a gimmick. in what scenario are you actually reading a 500k line diff?
something i'd really want to see from forges is alternate diff techniques: like AST diffing.
Performance and optimization is one of many pieces, but yes, it's a meme to render 500k lines.
That said though, and maybe I didn't say it well in the post, the more performant and optimized your tool is, the less burden you put on developers and users.
Sure you won't review 100k lines, but maybe the diff includes a ton of testing snapshots, or maybe it's a long running feature branch and you need to just quickly jump in and look at a specific change from a specific file. The less the developer or the user needs to think about `how` to render the diff or `how to navigate the diff`, the better we did our job.
Optimizing for the P95/99 case of performance typically makes everything better as a whole.
I don’t have much to say except I appreciate the ASCII art.
Is this really a hard problem? Rendering some text with coloring?
I believe we're just making everything harder than it needs by constraining it to a browser interface. This would be trivial as a terminal application written in native code.
Yeah, my bad. Sorry i wasted your time :(
What are you even talking about, this post is about rendering code diffs in the context of the web.
Whether or not it would be "trivial" in a native terminal application is irrelevant.