Patterns for Performance

♪ ♪ So, yeah, I want to talk to you about patterns for performance, because I guess we've already seen some talks today which have been talking about performance, and as a React developer, that is a key thing that a lot of us are always, I guess, important to us, important to our users. So that's why I want to share this with you today.

As I do work for Azure Grid, there's going to be some grids featured in this talk. So, can you say which of these grids is better? We've got the top one where you're resizing and it's, you know, jumping along, or you've got the bottom one where you've got this smooth resizing. So I think it's obvious which ones our users would want to use and which you would want to work with. And so this is something we're going to dive into and say how we solved this issue.

So, as you already know, I work at Azure Grid and our mission is to create the best JavaScript or React data table in the world. So, you know, we've got a free version. We've got enterprise versions. So do come and find out more about this because we think it's a brilliant product and it's getting better and we are really heavily investing in our React component. So we hope you can try it out, but that's enough about Azure Grid. You want to know about performance.

So to, I guess, give us a story for this talk, we want to have this reconstruction where a user has started using Azure Grid, very old version, which doesn't have this issue anymore, and they said, oh, you know, when I use this custom component, the grid does this really slow resizing. So they think, oh, okay, but they also say, well, it's when they're using a custom cell renderer. So just to give you, I guess, the background you need, custom cell renderers are a way for you as a React developer to give our grid a functional component or a class component and say render this within every cell. So here we've got a weather table where instead of numbers, you've got pictograms based on how many days of sunshine. And that would be your own React component. So in the use case, we've got they're displaying a total. So they might have a renderer like this, where in the body of their renderer, they've got some really complex logic, which is really slow. And then they're just returning a value. I mean, this isn't going to be slow, but just go with me and hopefully it'll demonstrate what we're trying to show here. And then to use this cell renderer, they pass it to the grid in this way. We'll find that on their column saying, well, actually render this with my renderer and pass it to the AG grid component. So we'll come back to this stage and we think, okay, how are we going to resolve this issue? Because we haven't noticed it yet. If we want to diagnose why it's slow, we need to just take a step back and think, well, in the normal case, when it works fast, how does the grid behave in that situation? So we take a benchmark. This is the standard grid, no cell renders, and it resizes really smoothly. We're going to use the React DevTools profiler and we'll profile it and we get something like this. So while we're resizing the column, we've got 254 renders going on. So that is a lot of renders, but they're really fast renders. So that's why you're not really seeing any slowdown despite the fact that there's that many renders.

So it's not always more renders is bad. It's also, well, how slow are those renders? Because React is fast. React can render a lot of times without your user and your application slowing down. So now let's go back to the user's case where they do have this issue, and we'll do the same thing. See, now we've got less renders, but each of these renderers, instead of being less than a millisecond, are now, what does it say, 28.9 milliseconds to render that total renderer. And each one of these is now blocking the main thread. So React, in this same time period where we're pulling the icon three size, is only able to render it 10 times. And so that's what's giving us this, you know, jump, jump, jump, because the browser isn't able to repaint enough times to keep up with it.

So why is it slow? Well, the cell component, it re-renders many times. And we saw that even without our custom renderer, that was still rendering a lot, but it was rendering fast. You introduce this total renderer, which is a child, so that's also being rendered a lot, but then it has an expensive render, which is now blocking the main thread. And so instead of all the fast renders, we've now got some slow ones, which is causing this behavior.

So why is it rendering so much? Well, so in the initial implementation, we had this, so the cell component had a state for the width. And so it was listened to this callback from our controller saying whenever the width changed, so when it's being dragged, it would call setWidth and set the state, and then React would re-render. So it's the standard kind of pattern of state re-render. And I think the talk two time before was, you know, also running in kind of these issues of how do we escape this pattern. Simple solution, Memo. You know, Memo lets us skip re-rendering the child component. So we can update our code like this. So instead of passing the value render directly in, we wrap it in Memo, pass it to the cell renderer, and that takes us back to the fast rendering. Now, if I stop there, that would be a very short talk, and I'm sure nothing new. But can we do better than Memo?

So let's profile it again, but now with the Chrome dev tools. So it seems like it's working fine. But if we take a closer look, when we're resizing, we'll see something like this. So there's lots and lots of these different actions where it's re-rendering as we're changing the side. And in total, it's about 81 milliseconds of scripting. And if we zoom in to each one of those spikes, what we can see is we've got AD grid code followed by some React code. And this is using the dev mode to give us slightly nicer, I guess, name. So we can see what React is doing. But what you can see is we've got AD grid doing something, it's painting, and then we've got React doing something. So this is, AD grid says this is what the width should be, and then we tell React, and then React takes that, re-renders, and does all this extra work. So Memo is still gonna be running comparison code. And the cell comp is still rendering. So even though we've stopped the child component re-rendering, which was the blocker, we're still actually doing potentially more work than we need to do. What if we don't re-render at all? And this is where it takes sometimes to look at our application code and think, well, what is actually changing here? So if we look at this line of code, the only thing that the width is changing is the style property. So we're not actually making any structural changes to the DOM. So when we see this, we think, well, actually, maybe I don't need this in a state. Maybe I can actually do this myself.

I can access the DOM, and I can change the style property. So how do we get the reference to the div? Well, you know, you can use a ref, and you can set it on the ref, and we know that that's gonna give us a reference to that DOM element, which we are then able to interact with.

So then, within our callback we can use this effect to set up it on a net, and we directly change the DOM style. So we get the sole ref, we get the current one, and then for that style property, we're now changing the width. And this is gonna have exactly the same impact on our component as what React would eventually do after it had rendered, worked out what had changed, and produced the new output. So we can just skip all of that, and say, well, actually, the only thing which is changing is this. I'm gonna surgically change it myself. And now, if we re-profile it again with this change, we're down to 31 milliseconds. So it's a massive drop, because we've just said, well, I know what I'm doing, let me do it myself, use the browser, you know, reference it in the way that we can do. So now, when we're resizing, we've just got ag-grid code, and the browser is repainting.

So I think this is something which we do need to remember. We're working in React, but we're also working in the browser. And the browser has tools and ways for us to update things that we can use to good effect. And so this is like one example of where we can do that. And so we've gone from this with react-memo into the direct-style-update, which is giving us a much better user experience. And also, something to note here is that it's not only a better user experience if the user has remembered to use memo, this is now also going to work if the user hasn't, the developer hasn't wrapped their component in memo. So we've enabled our grid to be much more robust to however it's being configured. And that's something which is important because even if developers are making suboptimal choices as an application library, we want to try and give them the best opportunity for our product still to work as fast as it can do.

But we need to be careful. The code which we've shown so far actually has a pretty big bug. I don't know if anyone has any idea what that might be, but we'll see this. So we made these changes, but then suddenly the columns start flickering into place. So you load up your grid, and you can see all of them are overlined over the first one. So what's happening now is that we're using an effect. And an effect is run asynchronously after the browser has painted. So what we've got here is React is rendering all our columns, painting them to the browser, and then it's running the effect, which is then coming in and changing the style directly, which is why we've got columns all in one place and then suddenly they jump. So we don't want this, because this is even worse. So how can we set this initial width then? Because you can't just put it in a state and then have your direct style manipulation, because the next time it renders, whatever was in the state value is gonna override what you've done. So you're gonna have to work out, well, which one should I be reading and what's the source of truth? But we've got these two alternative approaches. So the first is use layout effect and then the use ref as a callback.

So I love this. When I took this snapshot from the docs, it was like, oh, okay. So there's this pitfall. Use layout effect can hurt performance, but we're actually using it to improve the performance of our application. So you've got to kind of dig into, well, why could this hurt performance, but why in this use case is it actually helping us to have a better performance in our app? And so the key difference here is that use layout effect. It's like use effect, but it fires before the browser repaints the screen and it's synchronous. So any code you put in a use layout effect is going to block the browser painting. So this is why it can be a pitfall. So if you do have some expensive code, you know, put in there, it's going to delay the time that the browser gets to repaint. Don't know if I'm making that sense. Which is why most of the time it says use an effect because the side effects, they might be heavy. And so, yeah, you want the browser to paint and then you run those. You don't want to have to block that painting.

But so we make the simple change. We swap use effect for use layout effect. We don't have to change anything else and now this code is running synchronously. So the width is applied before the column is rendered. So we don't get this jump. So this fixes it. So that's a viable solution for it. But there's some potential downsides and that's the use layout effect. It's going to be run no matter whether this cell is mounted in the DOM or not. So it could be conditionally rendered but the logic in this use layout effect because it's dependent on the component itself, if I go back you see it's got this, the empty dependency array. So whenever this component is first rendered this code is going to run. So if there is something in there that's slightly expensive, it's going to be run every single time. And also it runs twice in strict mode which can then mean you have to take other actions. And there's a good blog post here from Dominic about using avoiding effects with callback refs. And so this is then a pattern which I think is an even better approach to solving this problem.

So we're all probably very used to having use ref in our code where you define the ref and then you pass it to the div element and then you know that that ref is going to be defined. But if we inspect the type of that closely there's also a callback. So we can actually give this a function. So we define a ref as usual because we still want to be able to reference our div element for the cell. But then we define a callback. So we're going to say set ref and this callback will receive a reference to the div or null. And you pass that callback to the ref. And the nice thing about this is that this callback is only going to get run when this specific div element is mounted. So if things are conditionally rendered this, the code that you set up in your callback for this reference is only going to run when the div is actually mounted. And then it, you can get null because React will call this again when that element is unmounted. Now it's crucial that you use, you wrap this in a useCallback because if you don't then on every render this React will recall it with a new ref. So the callback there is essential. So then if we come back and update our code for our cell reference, it will look something like this. We define our reference and we set up this callback. And now within this callback we do the same code where we listen to our cell controller for the width changes and apply the style directly. And this now works, runs synchronously to start with and then also enables us to update the width dynamically or directly without causing any renders. So I think I ran through these. You call it, ref with the elements be mounted, it's gonna be called again when it's unmounted, run synchronously which is key for us and it only runs once. So you can avoid some of the strict mode logic if you need to.

So the final result is, no matter what the user passes in or the developer passes in, we are now guarded against slow renders because we don't need to rerender when we're just changing the width of this column. So we've got two renders here instead of 254, which I think is a pretty big win. But some takeaways as I finish here. So React is fast. You don't have to prematurely optimize everything. So do wait until you have a performance issue before you start adding this complexity into your app. As you saw, memo is likely good enough, but with Agigrid we wanted to push it to the limit. We wanted to ensure that no matter how much data users put into our grid, we wanted to make sure the grid still performed because after all, that's why people are choosing to use Agigrid. So we needed to make it as fast as we could. And taking direct control, it can lead to these performance gains, but make sure you use the right tool as we looked at the different approaches to use the reference so that you don't end up introducing new bugs into your code. Thank you very much for listening. I hope that's helpful.

So, okay, let's start with this one. Some tech support, please. Our Agigrid license has spaces in it, something like my company. This prevents us from adding it in the .env file. We would like to hide it from Copilot. Okay, yeah. As in, you've got a license, so send a support ticket and we'll look into that for you. Nice one, on to the next.

Does this apply to React in general or only for Agigrid-specific use case, we don't use Agigrid? Yeah, as in the code that I've shown is just all React-specific. This approach can be used. So I guess the important thing will be to not take away from this talk... Well, I do want to say, yeah, Agigrid is great. But hopefully, it will also just give you that idea if I'm running into a performance issue, what are the approaches that I can take? So I guess this morning, there was about using the micro task queue. This is another approach where the key thing here is to say, well, what's changing when I'm updating my state? Is it something that you could take care of yourself with the browser APIs? And if you don't need to use what React gives us in terms of then working out what DOM elements have changed and using the virtual DOM, then that's something which you can look to control yourself.

I love that answer. Thank you so much for that. And then what happens if a custom cell needs internal recalculations based on width? Yeah. So then that logic then can be done using, again, browser APIs. I think there's ways of monitoring how wide your container is. I think it's mutation observer. There's ways of doing that. But this is a valid question because I guess what it's pointing towards is, well, what if the width was passed down as a prop? And so then if that was part of our interface to say, we give you the width, then we wouldn't be able to use this workaround. So this is where you are constrained by the code that you've got. And if the width was a property that we passed into these custom renderers, then yeah, we wouldn't be able to do this and we'd have to make sure that things were memorized correctly. But again, as the width is a prop, it would break, useMemo would. Or the memo would still render the child component. So for us, it was, well, actually, most custom cell renderers are written in a way that they are dynamic and they can resize within that so that we don't want to pass the width down and run into this issue.

John says, hi, great work with AG grid. Thanks, John. How do you compare to the TanStack table? Yeah, we love TanStack table.

And I think you might, if you look closely, we sponsor Tano because of the work they're doing. And we see ourselves as not competitors, but as in two different approaches to solving the same problem. So like for AG grid, we like to say it's batteries included. So you just drop our component into your page and it takes control of, and takes care of it for you, displaying all the rows, displaying the headers, the cells. You're just configuring it. Whereas with TanStack table, you've got to write all the UI components yourself. You've got to write your cells, your headers. It gives you all the building blocks as a headless table and the functionality to say which rows will be displayed but that's up to you. So yeah, TanStack table is a lot lighter because it's doing a lot less, but AG grid is a much more drop it into application, fully featured working out the box, no real work for you to do, especially as it's free as well, well free tier and an advanced one.

And do you have any book recommendations or courses to learn about patterns for performance? Well. Are you building one? I'm not building one. But as I was reading, I just heard someone on Twitter yesterday who has written an advanced book. And so they, I would have to find it. I don't know their name off the top of my head. You may already know them. But yeah, there's definitely courses out there. I can't remember though. Okay. Sorry. That's fine. And here we are.

Okay. Is this a pattern you use often or is this a very specific example for this use case? I think it's about say this is a specific example for this use case. But again, it comes back to that point of spotting these patterns. So it's like if you run into this in a different situation where you might see a similar thing is happening, then you can apply this pattern again. So yeah, it might feel very specific to that use case but I think the important thing to take away is there's a pattern here that we can then apply if we run into this situation again.

And are side effects in a set callback essentially a hack? Side effects in a set ref callback. Let me think about that. We can come back to it. Yeah. I said it comes to what I mean by side effects. I guess there's anything which you are... I guess you got to be aware of the way that these callbacks are running and that they are the synchronous one, so there's some things in there which you might want to avoid. But I guess, you know, whether it's a hack... I don't know, you know, a lot of the things we do, some people will say, well, that's not the standard pattern, so that's a hack. But I think when you're pushing the limits, you sometimes do need to step outside of the standard path. So, but then that's also where, maybe don't reach for these initially, only go for it when you need it.

And what happens if you need to persist the column width for next runs of the application or in a tabbed app? Well, that's a great question. I think I know who's asking that question.

So in our next version, in version 31, and which is coming out end of November, it seems like they've got some inside information. We're gonna be adding a new feature, which is like initial width, where, when you tab away from your grid or you destroy it, you can get the state of the grid. And then when you tab back, you pass that same object into the grid and it will restore the grid to the same width, you know, same column width, same order, and remembers all that state for you. And that's something which we know people have been doing. And so yeah, in 31, which is coming out end of November, that feature is gonna be implemented. Very good question, and it leaves us with something to look forward to as well.