Co-engineering the future: How Uber & Cloudflare engineer a global platform
Steve Goldsmith, Mattie Toia
Cloudflare · Uber
Canva is empowering the world to design. With over 260 million monthly active platform users and more than 400 designs created per second globally, Canva relies on trusted partners like Cloudflare to deliver solutions today and co-engineer the future. In this fireside chat for platform architects and engineering leaders, Cloudflare VP of Product Steve Goldsmith, Canva Senior Engineering Manager Michael Yates, and Canva Principal Software Engineer Adam “Laz” Lazur share examples from their joint engineering efforts, with detailed technical explanations and anecdotes from real-world delivery experience at Canva. Leave with concrete insights on balancing security, scale, and high-performance global platform tuning.
0:02Thanks, everybody. I'm Steve Goldsmith, VP of Product here at Cloudflare. These are the brightest lights I've been in front of the whole time. So it's not you, it's me. This is the last of the breakout sessions. And then we have our last general session, which is really cool. And then it's beer 30. So I'm excited to kind of wrap this up over is that great couple of days. I'm joined here by Michael and Laz from Canva. We'll be talking today about just sort of how Canva scales on Cloudflare. Real frank, technically. discussion about how this company who obviously everybody knows all of us use it is and we're a small part of how they're successful and we really like to get into the weeds of technology and how they adopt it do you guys want to do a few minutes of introduction about what you do at canva yeah sure uh my name is michael yates i'm a director of engineering at canva and i run a section of canva called api platform which means we look after all of the traffic engineering north south and east west uh i'm adam lazar by Laz.
0:56I'm a principal engineer at Canva and I work with Michael's teams to do all the traffic things at Canva and then solve other engineering problems, I guess. Excellent, thank you. So we're going to start a little bit of a backward-looking history lesson, sort of like how big was Canva when you started and what was the adoption of Cloudflare, which products did you start with and also just throw out the question and see where you want to take it. Yeah, I can give a little bit of a story here, because I... thought this might come up um canva has been using cloudflare for more than 10 years it's actually 10 and a half years um since we sent our first bill or check um to uh to cloudflare and i looked up the document about the decision that was made to choose cloudflare and it was for uh cost reasons so they wanted to move off aws that they were using at the time to front a bunch of resources.
1:56And the growth of Canva through that time has loosely mirrored the growth of Cloudflare during that time. So the first enterprise deal that Canva signed on to was for $7,000 a month. We pay a lot more than that now. Canva at the time had around $2.5 million to $3 million. million monthly active users we now have a quarter of a billion so 250 million so roughly over that 10 years 100x so that's incredible yeah and in terms of the products you started with what was that it's like cdn was where you started yeah from from looking at that historic document uh it was absolutely cdn yeah okay and then in terms of adoption since then you sort of added you have the whole menu from us pretty much at this point so how what's the rollout or adoption sort of along the way it felt like.
2:49Yeah, so from what we can work out, and again, a little bit of enterprise archeology to kind of work that out. There's broadly two main sections where we use Cloudflare. So we work a lot on the traffic side. So camera.com coming in and the product. We also have a pretty big enterprise footprint with you. So we use warp for our end user compute. We use... cloudflare access to protect a whole bunch of internal applications um pretty much every product that cloudflare have we have one of in some form so um the adoption has followed the growth of canva so we try out a lot of stuff um it doesn't always work but first off but yeah we try out pretty much everything and then we make decisions about whether we want to continue with that or not so well it might be helpful to kind of explain sort life of a request in Canva and how that moves through your platforms, our stuff, and other providers.
3:48So sort of starting with like, I'm a user, I open up Canva and I start to do things. Can you maybe both of you just kind of walk us through like how that flows through different stacks of technologies? And this is a technical audience, so we're going to get into the weeds. 100%. Well, we can dive right into the weeds with Laz. I'll probably hand over some of that to you. Sure, sure. So Cloudflare, our edge is where users initially connect. Our product is a React -based web application, primarily. And so users fetch assets from the edge. They start to load our editor or our home page. Dynamically, they're fetching assets. Then they'll make API calls through the edge. Behind that, we have a single origin.
4:35Our architecture is actually pretty simple. You hit this single origin for a variety of different things. And then we route to many, many, many microservices across the stack. The other interesting aspect of our application is we have a WebSocket that's persistent that our clients use to synchronize edits to documents and do some more sophisticated things. And then we also have synchronization of mouse pointers on shared documents and things that makes us a little bit unique. So there's three traffic types there, right? So you've got your HTTP requests coming in. So we open up a WebSocket for real -time requests and then WebRTC for collaboration. So three main traffic types. And then within each of those, we sort of break it down a little bit further.
5:23So on HTTP requests, you know, telemetry comes over HTTP requests. Not particularly interesting traffic. Our APIs come in over HTTP requests. Much more interesting to us. Obviously, it has to be... Ah. low latency, high performance if you're doing APIs. Just broadly, in terms of architecture, I think the other thing you didn't touch on is that initial JavaScript bundle that goes back up has the first lot of hydration, and then after that, the APIs are used to hydrate the rest of the space. Yeah, we do the bootstrap, which is like a big, it's an optimization where you send one big payload that starts the editor running, right? optimize the kind of time to interact them on our editor or home page. So we do some kind of aggregation of multiple API requests that we do at standard startup and we do that in one big burst and try to get it out to the user as quickly as possible.
6:16We're extremely focused on getting the user to using the product as quickly as possible. So down the HTTP request life, every request that comes into Canva transfer. a worker and so that means for us we process trillions of HTTP requests a month inside that worker we run a router and so depending on what we find when we open up the HTTP request so hopefully give us the HTTP request we've got access to the full to the full request we then run a routing layer inside the worker which steers traffic to different parts of our origin so in broad terms you can think about apis coming to our gateway and you can think about assets going to s3 the the routing table is much more complicated than that um it has around 1500 entries i think it's a lot um in a tri structure if anybody remembers yeah cs201 uh cs2 yeah um it's the same in australia cs2 um so if you're gonna try uh and yeah workers, I think the P50 is around 2 milliseconds, between 2 and 3 milliseconds.
7:33So three different traffic types come into the Cloudflare Edge and then go through this workers-based steering component and then hit your origin. We can keep going down. Yeah, let's keep going. Once it hits the origin, then how does it work? So at the origin, we run a couple of different pools of layer 7 load balancers that route to microservices behind them. Broadly. we've moved to an Envoy -based architecture with our logic kind of sidecar setup. Historically, we had our own gateway that was written in -house, and we moved to open source because it works really well. Behind that, there's a lot. We have a massive microservices architecture. I think the Envoy configuration will route to maybe 250 or 300 different microservices, and then behind those, there are even more things.
8:26So at the gateway, so we come in from the edge, we come into the gateway, the gateway's envoy now. We look aside for some auth, which we've been doing for a long time. We do some protocol translation inside the network where we're gRPC, outside the network where HTTP. So the gateway does that protocol translation for us, does some work around streaming to convert into gRPC.
8:58And then on the way back, obviously, send the request for that. So where do you feel like you're working around versus working with Cloudflare in that flow of transactions? And I say this knowing that you have some of the highest scale and a single domain and lots of things that make Canva very special in terms of Cloudflare's customer base. special in a unique way. Everybody's special. I mean, there's a lot. So one odd and, well, I guess interesting thing about Canva is we run just about everything on www.canva.com. And so we have some great benefits from doing that. But also, there's massive problems with cookie payloads. And we do lots of progressive routing. And so tuning things there is an interesting aspect where sometimes we work well together, and other times Cookies are expensive and problematic.
9:53We also use the WAF to protect us from a variety of things, and it comes with cookie payloads, which is good and bad. So we've been doing optimizations around that. So that's one aspect of it, I think, where there are good and bad aspects of the site. I'm trying to think of other... I mean, there's a set of stories here as... Canva has evolved, we've asked for various things to evolve on the Cloudflare side. So an example of that, we had a worker and all traffic was running through the worker. So I've been at Canva about three years. And when I arrived, the P50 was much higher. We were doing more work in that main worker and the code wasn't as efficient.
10:42We couldn't re-architect that worker. because of how the workers' billing model was. So this is the interaction between the software side and the Cloudflare platform. And this story's got a good ending, so it's okay for Cloudflare. We spoke to you about that. We now have been able to use workers and sub-requests. We've been able to re-architect our worker and bring that router to the... to be part of the main worker and other parts of the work that we need to do are happening in sub-requests. And what that has allowed us to do is run that router on the edge, but also we can run it offline and process all the HTTP access logs and do route attribution for cost inside the organisation.
11:38So we are still working around... fact that at our scale it's important to know historic trends of which routes we hit and how much they cost us to run but you know we continue to evolve the platform to help us solve some of these business problems so it was like public facing routes or internal like microservice public facing public facing just so what we've tried to achieve with the programmable edge is is to be able to scale what teams need to do and give them safe config in order to be able to configure their route how they need to. So we've unified our configuration and we've given each team inside Canva, if you want to think of each team owning a route, we've given each team in Canva the ability to configure those routes.
12:34We then bundle all that up and ship it into the worker. um other challenges we've had with that model uh up until relatively recently workers didn't have progressive deployment so we had to work around that we ran two workers and had to do some interesting tricks for traffic splitting so that we could a b um our two versions of the worker um testing is difficult you know so there's been a bunch of different um challenges you the platform keeps evolving as we evolve so yeah we're happy with that that's good um you know this is sort of a follow-up to that was there a specific spot along that curve where you feel like the relationship with cloudflare changed it was not like you know a component that you had picked but it was more of a strategic partner that we needed to start working more closely with and and thinking about like a longer horizon of working together so an interesting question for me because i arrived three years ago and my sense of things at that point was that canva was uh very much treating cloudflare more like a vendor and less like a partner um our account managers here megan um we've worked hard to make sure that that's not the case yeah um and so we as we start building things we consider things from the cloudflare menu um that menu is getting bigger every day so you know there's 150 things to choose from now um and that
14:07that menu is quite large but i think i don't want to kind of say when i arrived it changed but it feels um it it feels like it's changed um quite dramatically in in the last kind of two or three years that's great so if we go back to kind of life cycle for question these different workloads are going through there how do you think differently about if at all latency, you know, global distribution of traffic, do those have different, you know, thresholds for you? Or is it all sort of like, it's all one client experience? You think about those streams in a similar way? I mean, I think we're hyper focused on our users and our users experience. So I've been trying very hard to tie our per request metrics into product level metrics and optimize the things partially because the business looks at the product level metrics.
14:58and they want to make those better. And when I say I made a request, P99, I go down by 5%, they're like, what do the users see, right? So we don't have strict micro -latency kind of targets for things. We have some guidelines, but we mostly look at the user-level metrics. And as we get more sophisticated, we're also looking at user-level metrics with distributions, right? The P50 user experience versus the P99 user experience is vastly different. i've been advocating very strongly for bringing the p99 way down right um that takes more sophistication but but again like we want to make it the best experience for our users and make it as fast as possible um and uh the p50 user has a pretty okay time but the p99 user is not super happy right um in that direction this is a dynamic product that people do creative stuff in it how do you think about caching how do you use our cache with something that's so fluid and dynamic. We try to cash everything we can, right?
15:54Surprisingly, so I come from a media background before Canva and cash hit rate's hugely important, obviously, for that. Our cash hit rate is actually 70%, which I think for a product, what's 68 % or 69%, for a product that is highly individual, that's a pretty magic number to be able to hit those kind of cash hit rates. One of the good news stories with Cloudflare is we don't have to think too much in the nicest way. We don't get charged for how much we cache, right? And so we do set cache control headers on traffic as appropriate.
16:39The Cloudflare Edge handles it for us. We don't think a lot about whether we want to or not. We just set the cache control header and hand it over to you guys. in the in the response and off it goes when you start to break down different asset classes or different traffic types you can start to see some really wild differences in how things are cached fonts as an example has something like a 99 % cache hit rate and then there's javascript maybe more like 25 % so you get some some very different numbers and then you have to What is the user actually doing and should we be tuning that up or down? We've actually just done some work the last two weeks or so where it'll actually be net beneficial for Cloudflare because we've changed how our JavaScript is broken down to provide more stable caching of the JavaScript up until a while ago.
17:46week or two ago every time we did a deploy the the fingerprint of the javascript would change for all of the javascript which on the cloudflare network you'd actually be able to see us when we did a deploy because everything got evicted and got refilled across uh every node uh that won't be or it's not the case now since uh since our changes so yeah um in in that sort of request thinking, what parts of that end-to-end experience are synchronously blocking through Cloudflare? I think you might have covered some of this. And what parts are you able to put off of the hotpath and come back to either post-processing or async
18:31We have some async API requests and callbacks. A lot of that stuff happens over the WebSocket because we can do whatever we want there. I'm trying to think of, like, what We plumb up chunked uploads and those kind of things. And often, they'll come with a transcode kind of component that comes back.
18:52There's a lot with the API requests. So actually, amusingly, most of our API traffic is limited to 30 seconds. And so it forces applications into that footprint. And then people push around it in different exciting ways. How do you think about? performance measurement in such a complicated like is it fast enough you said p50 p99 for certain components but like just that end to end when you think about answering that question of what's the user experience that's a hard one to answer it's really difficult yeah yeah um because so canvas mission is to empower the world to design we we take that pretty literally um and so we can see massive differences in performance in uh different locales um and so So sometimes it's the ISP into Cloudflare.
19:43Sometimes it's Cloudflare's backhaul network. Sometimes it's the interconnect between Cloudflare and the ISP. Quite often it's our code, like we're not perfect by any means. So figuring out those latencies is kind of a full -time job. So how do we think about it? I think at a really high level, we get as much telemetry out of Cloudflare. as we can and we overlay that against our our user metrics um i was running some numbers recently where we spend a lot of our bandwidth on cloudflare on on telemetry of various kinds so we bring all that back into a big data warehouse so we we look at that uh and we try and make it as fast as we can uh from from a network point of view and then we hand off some of that our application engineers and say hey like this bit's not you know it's too fat you need to slim down the amount of javascript or you know provide stable hashes as an example to speed up the application i mean we into we we do the low level bits and bytes but mapping into the product and the usage like we an example is we're trying to compress as much as we can to send fewer bytes on the network right um but i also talk to product people about it's better to send no bytes than to have to compress bytes, right?
21:03So if we can reduce the payload or reduce duplication, there are lots of things that we can do in the product. And I'm happy to do compression, but there are also product things that we need to push back on. Telemetry is another example where we aggregate a lot of telemetry. We were batching inefficiently. We weren't compressing. There are lots of ways to optimize the code to make the network's job easier, right? And sometimes the problem is that we have to use the code to make the network's job easier, right? people don't understand because the product works okay and i'm like i mean one fewer bite on the network is one fewer opportunity to drop a packet to have problems right so like we it's better for everyone if we send less data right if we have to send it we should but um so we i rove around the stack advocating on behalf of like we can make this so much better like there's so many things to do we have tons of data um and tons of ways of analyzing it uh i try to line up with the standard product metrics that people care about because because that's how the business runs um but then also try to pivot to like look at these latencies look at this distribution uh often people like to make decisions where they're like oh new feature x it like makes everything better but like often it's like well it makes this better but that worse and the decision is gray and so i try to engage with people get metrics that we understand that we believe in true because you can also completely deceive yourself with metrics very easily and just work the entire system.
22:34Everybody has to collaborate to make this better. But it's a lot of translation, too, between the different layers and getting the people at the right context to do the right work. I know from working with you that Canvas started as a web app, but your push is more focused on mobile end user use cases going forward. How much of that changes when you start thinking about you know thick client and mobile client versus the react based stuff yeah i mean it's an active area for us especially on mobile we've got teams now who are looking at rewriting things that traditionally would have been done in javascript inside the web view closer to the metal if you like we've got people who are looking at redoing various components most of that looks like it's Rust, and then we've got a binding layer with FFI, if anybody's familiar with that, to start binding in some of those high-performance components or higher-performance components into the mobile app.
23:39So it's an area of active investigation for us, and we're looking at which parts of the mobile stack, as an example, make sense to move in. those lower level languages and still more to come and you know we haven't released an app with that yet but if you think about something like video how much video can you do on on a client and then we start to then think about well that's great if you've got a two thousand dollar iphone but if you've got a hundred dollar android device in
24:19Nigeria or Philippines or something like that you're probably not going to be offloading work to the GPU and that's where we can start if we write some of these low-level components in a systems language we can start to have conversations about running it okay well it's not on the device maybe on the edge and then if not on the edge maybe in core compute and so we're thinking now about because scale and speed are important we're thinking about portable portable libraries and portable constructs from our edge to our gateway and our core. That's a really good pattern. That makes a lot of sense. And I admire the, it would be easy to focus on the $2,000 phones and it's a much easier problem to solve, right? To actually solve for the world's creators makes you solve the harder part of the curve. That's, yeah, that's our mission.
25:07And it's some pretty cool engineering when you get involved in it at scale. It's good. So when, speaking of building new things, building a new product you have a new idea how do you how and when does cloudflare come into that and sort of the follow -up is do your development teams know they're using cloudflare or are they using like an abstraction or library or whatever that's sort of like or a little bit of both depends on the use case i'll set up um i'll set up this and then i might head over to las who develops every day um canvara is about 5 000 employees about two and a half thousand of them where some type of technology had. So we're a pretty big development organisation.
25:52Cloudflare is one of a number of, you know, SaaS vendors that we use, and the way I think about it is they are part of our toolbox, so they are definitely on the paved path. I'll hand over to Laz, but we have an infrastructure... abstraction layer that we've written in -house where we can abstract away some of the complexity of deploying onto various clouds and deploying into various places. It's called Platy. For those who know Canva or know any Canva employees, you can have discussions with them on Platy. Platy's short for platypus, which is an Australian animal. Nobody needs to know that piece. You're all here for the technical piece, but I'll hand over to Laz about.
26:40what constructs we abstract and and how how users use them internal users i think we have we try to get engineers to build using the most appropriate platform and tool choice is always part of this right um and depending on where we are in the life cycle if it's a crazy new idea that's driven by a product that we don't know will be successful we will try to launch it as quickly as possible uh to get feedback on it right get it to our users um If it's something that we know will be a strategic longer-term vision, then we'll pick different platform trade-offs. An example of a recent note is we launched the Canva MCP on Cloudflare because it allowed us to launch it as fast as possible. Instead of building that technology in-house, we could utilize a Cloudflare abstraction and just go, right?
27:26It was fantastic, I think, for everyone involved. In other places, if it's a bigger investment or we can't... buy it from a platform, and then we'll in-house it. We're doing the build versus buy conversations constantly. We do have abstractions. And often, many of our product engineers, they know Cloudflare, but they don't know how the caches work or any of that stuff. They don't have to operate at that level. But certainly within the infrastructure, everyone knows about Cloudflare. And then once things take off, you have an idea, you rapid prototype it, you launch quickly, we pay some money for our platform or whatever, and then it- catches on and it's large scale and it's expensive or slow or then we optimize right we double back six months a year later on that product and we change it so that it's efficient to serve it's faster better stronger on the product side they're always innovating on the product stuff but we're also our infrastructure is refactoring itself and moving workloads around the network for the most appropriate place and sometimes that's performance sometimes it's cost sometimes it's availability A big kind of macro theme and infrastructure at Canva is being flexible as we continue to figure out how to do our product better, faster, stronger.
28:39The world's also changing around us, right? So we're adapting to new AI demands. We have to have GPUs over here. And as that pops up, the engineering organization that serves the product has to adapt to these things, right? And so more and more so today, we're thinking about out not one platform not one product but how can we bundle a bunch of these things together with abstractions so that i can be flexible about where i run my workload um how things find it how we store state how we distribute state um and to do the computation the storage the communication at the right spot in the stack but again early product like launch it figure out how it works and then once it's once it's caught on then we iterate, and refine, and get good at it at a larger scale.
29:32Often, the product people, I love product people, but they'll come and they'll be like, they're like, Sit right here, man. They're like, this is a great new feature. It's going to launch. The whole world's going to use it. He's in rainbows, and tell me about it. And I love their enthusiasm, and I'm like, I mean, I've seen this a million times. But we enable them to launch in small-scale ways, to get feedback early. do like rapid, get that in front of people and figure out how it works. And then if it's a success, then we enable them to do it at scale, right? But the important part is figuring this out and getting the technology out, and then iterating, iterating, iterating. So speaking of new stuff, then there's a couple of things we've been working on together that are your, as I mentioned earlier, helping push our platform forward, because Canvas scale always helps us do this.
30:20Can we talk a little bit about some of the two or three projects that we have in flight with you? that partnership's working and sort of like areas where we're collaborating together. Want me to start? Sure. We've spent the best part of the last 12 months re -plumbing a whole bunch of our north -south traffic infrastructure. We touched on some of it just before with Envoy and the place we're going next is actually very much blurring the lines between where Cloudflare stopped. and where Canva starts. So from a network point of view, we're trying to work out a control plane where we'll be able to control the distribution of traffic directly in Cloudflare down into Canva protocols, XDS.
31:10And we spent a long time, or not a long time, we spent the last weeks, months, working with Cloudflare on that. And it's looking, shaping up really exciting. Laz has been deep in the detail of this, so I'll hand it over to him. I mean, it's the flexibility strategy. So a year and a half ago, I started at Canva, and we were deploying service mesh and trying to figure out how microservices could find other microservices.
31:40And the routing of that traffic and figuring out how that whole system works turns out to be pretty hard. And as we kind of iterated... it and we're figuring things out we stumbled upon xds and envoy speaks xds grpc speaks xds and so we with our microservices had a strategy for routing and requests we're using envoy as our gateway so we can also route requests there and so we decided we would kind of heavily invest in this as a strategy and so the next step out is we have the edged origin hop that we tell cloudflare we're like hey here's our dns name for our origin and you pull it and we're like well, you could talk XDS, right? And so we've been working together to develop this so that we can use this flexible control plane protocol that in its simplest form is just a replacement for DNS, like here's the IP addresses, but it also comes with here's how to talk to me, here's how to retry, here's how to route this left, right.
32:36And as we build it up, I think it will be important strategically, but it's a core part of our strategy is as our resources move around the network. we get gpus and a neocloud we need a way to route request to them and or levering these open standard protocols and plugging components into them um and cloudflare is a huge part of this um yes that that's one that's very network heavy i think um higher up the stack um there's an internal tool that we're um developing at the moment uh for for canvas staff based on real time um as well So we're not 100 % sure whether that will be a product that goes into the main part of Canva.
33:21But 5 ,000 staff is a relatively large footprint and we've got a bunch of software developers. So we're looking at moving some things onto some tools that we're developing in collaboration with Cloudflare. I can't talk too much about it because it's not necessarily... going to make the public Canva product, but it's an example higher up the stack as well where we're taking primitives from Cloudflare and, you know, stitching them together in some interesting ways. So there's many stories like that, but that's one that's higher up the application stack. I know, like I said, this is exciting for us because, you know, the scale that you operate at, the regions, the ASNs, the networks, the devices, every aspect of Canva.
34:09pushes us. Even the fact that it's all one domain pushes us to think about our product differently. So I really appreciate it. I got one more question, then we have just a few minutes for audience questions. If you have something in mind, get excited about it. And we'll get a microphone to you in a minute here. If you could get one thing out of the Cloudflare team, I mean, you have a straight line to a lot of people, so this is not really hypothetical. But if you get one more thing out of the Cloudflare team, what would it be?
34:36Well, I I've got my wish already. You can think. I'll think. So I've been doing a bunch of, I'm like a protocols guy, right? I've been doing a bunch of work trying to optimize things. And so the one wish that I would have today would be to tune the H-Pack buffer size up. HTV2 has this compression thing for headers. You get 4K that's shared between the client and the server. We have 6K of cookies. And so that. 4k is useless we just thrash that cache on every request and so we're sending an extra 6k essentially with every request to our massive single domain i can't get them to delete cookies on so if we could get it to be 8k or 16k and i could fit it in there then every request would be smaller and so that would be my one wish but then i'm also rolling out http3 right now and that one has the the shared header uh Dictionary, it defaults to zero and Cloudflare sets it to zero.
35:38And so I would like it to be on zero for similar reasons, right? Like it's about performance and optimization, right?
35:49We have some capacity constraints limited by the physical world. There are places in the world where it's really, really hard to get a good bandwidth. And there's a chicken and an egg problem. because good bandwidth can often lead to GDP uplifts and a whole bunch of other things in kind of some countries. I would love Cloudflare to build out more capacity in some of the places where we know we have problems. And I don't want to point to any specific markets or regions, but ahead of us being able to expand into those regions. So when we talk about empowering the world to design, we mean everybody. And if you can't get a good connection for a cheap price, that really inhibits your ability to use the product.
36:37So we're working as fast as we can to reduce the weight of the product. On the capacity side, there's some places that would be great to have a bit more capacity. Both very high level and low impact and network infrastructure. Okay, we have a few minutes for questions. Anybody have a question in the audience before we wrap it up? We've got one in the front here, please. We've got a microphone coming, so just wait one second. Thank you. At such a large scale, have you guys been able to utilize Cloudflare's durable objects when you have multiple users on this same application You want to talk about it? I kind of don't. You can talk about it. We are not using durable objects.
37:23We have been able to do this. a deployment in china there's canva .cn and durable objects are not available in china so we can't lever it there's some there's some small use cases where we do use um do for for a couple of events and things um the other part about camber just to know for the room um we own a set of companies so camber .com is by far the largest amount of traffic but um we own a company called pexels a company called pixabay a couple other companies affinity and some of those have some small deployments of durable objects but on the main product we're not a big durable objects consumer we've played with it yeah we know what it does yeah it is an example though of how you push our platform that we have a china network product that lets us work your partner and deploy you know canva.cn for a customer like this the menu of cloudflare does not all automatically work in that secondary china-based deployment and so It's another vector that we work with y'all on and say, yes, it works in China, but we need this and this and this.
38:27It's a complexity for us to manage. It's a tricky situation. And we try to keep it as homogenous as possible, but it's difficult. One last question. Alex Krivet.
38:40Hi. My question was, you had mentioned both, you don't want to design things for just the users with the $2,000. devices at the same time canva has such a wide breadth sometimes you're operating in really lossy networks across the world how as a company does how does does canva approach that from an engineering perspective how do you build empathy inside your engineering org to make sure that you're you're building for those people who are probably in well connected in well connected devices when they're building but maybe deploying to less well -connected places um i've got a recent example of this um you can actually go inside our resource booking tool and book a low -end device so we've actually got a little in in sydney there's a room um that's been set up with a bunch of relatively inexpensive or low powered uh low -powered android devices um the the wi -fi access point in that particular place uh also has some uh some clipping done to reduce the bandwidth um and we got contacted i didn't do my work and actually gave it to las um we actually pulled down some stats from the cloudflare network and gave them to our it um our it team who provisioned up like a p50 uh bandwidth um so that when you go and pick up that low power device you're getting close to um the
40:14the performance of what that person might see. That's an example of where, as an engineer, if you're working on a particular feature, you can go and grab that device and play with it. When we present things internally, often they will show a video of a lower-end device. It's a high focus for us right now. The other thing that I like to do, because I'm a network bits and bytes kind of guy, is I explain to the product people that every byte we send costs money, and they're like, I mean, do you want to cost... reduce our bandwidth and I'm like yes but also some of our users pay per byte right and every byte you charge them like is is an aggregate bad right so if we could all save money if we send fewer bytes um and I guess people in the U .S. I have a flat rate plan it's unlimited etc etc so you don't really think about this stuff but in other places bandwidth is very expensive and if we're wasteful with it it costs people legitimate lots of money right and and uh I try right like you you have to understand this is this is non-zero amount of money for people we should be efficient um and we're when we're building the application we're starting to do things where you kind of adapt with the device or technology so we'll build the you know high pollutant thing for the iphone and then we will scale it way down for android to be very efficient right um try to give similar product experiences but we didn't actually do this but i did threaten at one point in our non-production environment and because we run so much traffic through the worker i did threaten just to put in a sleep yeah and and basically if people were testing in
41:48in non -prod um and then and then tune that number so like this week we're operating as if we're in the philippines and next week we're operating as if we're in nigeria so those toggles are available to us but um at the moment the way that empathy is done is through education and a little bit of device testing. Plug for Ryan's Beacon product over here. We just announced the birthday week.
42:16It's web metrics that are not tied to Chrome that are available for your data warehouse type lookup because they're in a giant BigQuery table. So you can actually get benchmarks and baselines anonymized and have your product compared to those things. So if it's slow in this entire country or you're just slow in this country, and that's the sort of thing that this public data set... us at birthday week one of the interesting hallway conversations we've had since we've been here is take everything that you know about uh rum you know and measuring and then go and what about the agents how are we going to think about that same uh accessibility question around yep
42:55um but but the the point is the same which is you want to provide a good experience so start with I'm trying to create the empathy and try and measure what you can nice we did make it almost all the way through the whole talk without saying agents I'm glad you got that in it's mandatory I think it's mandatory it's mandatory we can't do a talk of this we talked about MCP though so okay indirectly directly well that's all time we have thanks everybody for being here really appreciate it thanks Michael and Laz for participating
Steve Goldsmith, Mattie Toia
Cloudflare · Uber
Jonas Bengtson
Discord
James Snell
Cloudflare
Be the first to know when registration opens and the agenda goes live.
--- title: "Enabling creators worldwide: Scaling Canva on Cloudflare" url: "https://www.cloudflare.com/connect/2026/session/enabling-creators-worldwide-scaling-canva-on-cloudflare" year: 2026 --- # Enabling creators worldwide: Scaling Canva on Cloudflare [Watch session](https://www.cloudflare.com/connect/2026/session/enabling-creators-worldwide-scaling-canva-on-cloudflare) ## Session details - **Track:** Web Performance & Frameworks - **Topics:** Customer Experience, Cloudflare product innovations, Performance & resiliency - **Solution areas:** Application Performance, Core Platform ## About this session Canva is empowering the world to design. With over 260 million monthly active platform users and more than 400 designs created per second globally, Canva relies on trusted partners like Cloudflare to deliver solutions today and co-engineer the future. In this fireside chat for platform architects and engineering leaders, Cloudflare VP of Product Steve Goldsmith, Canva Senior Engineering Manager Michael Yates, and Canva Principal Software Engineer Adam “Laz” Lazur share examples from their joint engineering efforts, with detailed technical explanations and anecdotes from real-world delivery experience at Canva. Leave with concrete insights on balancing security, scale, and high-performance global platform tuning. ## Speakers - [Adam Lazur](https://www.cloudflare.com/connect/speakers/2503407) — Principal Engineer, Canva - [Steve Goldsmith](https://www.cloudflare.com/connect/speakers/2503409) — VP of Product, Performance & Infrastructure, Cloudflare - [Michael Yates](https://www.cloudflare.com/connect/speakers/2503413) — Engineering Director, Canva ## Transcript [0:02] Thanks, everybody. I'm Steve Goldsmith, VP of Product here at Cloudflare. These are the brightest lights I've been in front of the whole time. So it's not you, it's me. This is the last of the breakout sessions. And then we have our last general session, which is really cool. And then it's beer 30. So I'm excited to kind of wrap this up over is that great couple of days. I'm joined here by Michael and Laz from Canva. We'll be talking today about just sort of how Canva scales on Cloudflare. Real frank, technically. discussion about how this company who obviously everybody knows all of us use it is and we're a small part of how they're successful and we really like to get into the weeds of technology and how they adopt it do you guys want to do a few minutes of introduction about what you do at canva yeah sure uh my name is michael yates i'm a director of engineering at canva and i run a section of canva called api platform which means we look after all of the traffic engineering north south and east west uh i'm adam lazar by Laz. [0:56] I'm a principal engineer at Canva and I work with Michael's teams to do all the traffic things at Canva and then solve other engineering problems, I guess. Excellent, thank you. So we're going to start a little bit of a backward-looking history lesson, sort of like how big was Canva when you started and what was the adoption of Cloudflare, which products did you start with and also just throw out the question and see where you want to take it. Yeah, I can give a little bit of a story here, because I... thought this might come up um canva has been using cloudflare for more than 10 years it's actually 10 and a half years um since we sent our first bill or check um to uh to cloudflare and i looked up the document about the decision that was made to choose cloudflare and it was for uh cost reasons so they wanted to move off aws that they were using at the time to front a bunch of resources. [1:56] And the growth of Canva through that time has loosely mirrored the growth of Cloudflare during that time. So the first enterprise deal that Canva signed on to was for $7,000 a month. We pay a lot more than that now. Canva at the time had around $2.5 million to $3 million. million monthly active users we now have a quarter of a billion so 250 million so roughly over that 10 years 100x so that's incredible yeah and in terms of the products you started with what was that it's like cdn was where you started yeah from from looking at that historic document uh it was absolutely cdn yeah okay and then in terms of adoption since then you sort of added you have the whole menu from us pretty much at this point so how what's the rollout or adoption sort of along the way it felt like. [2:49] Yeah, so from what we can work out, and again, a little bit of enterprise archeology to kind of work that out. There's broadly two main sections where we use Cloudflare. So we work a lot on the traffic side. So camera.com coming in and the product. We also have a pretty big enterprise footprint with you. So we use warp for our end user compute. We use... cloudflare access to protect a whole bunch of internal applications um pretty much every product that cloudflare have we have one of in some form so um the adoption has followed the growth of canva so we try out a lot of stuff um it doesn't always work but first off but yeah we try out pretty much everything and then we make decisions about whether we want to continue with that or not so well it might be helpful to kind of explain sort life of a request in Canva and how that moves through your platforms, our stuff, and other providers. [3:48] So sort of starting with like, I'm a user, I open up Canva and I start to do things. Can you maybe both of you just kind of walk us through like how that flows through different stacks of technologies? And this is a technical audience, so we're going to get into the weeds. 100%. Well, we can dive right into the weeds with Laz. I'll probably hand over some of that to you. Sure, sure. So Cloudflare, our edge is where users initially connect. Our product is a React -based web application, primarily. And so users fetch assets from the edge. They start to load our editor or our home page. Dynamically, they're fetching assets. Then they'll make API calls through the edge. Behind that, we have a single origin. [4:35] Our architecture is actually pretty simple. You hit this single origin for a variety of different things. And then we route to many, many, many microservices across the stack. The other interesting aspect of our application is we have a WebSocket that's persistent that our clients use to synchronize edits to documents and do some more sophisticated things. And then we also have synchronization of mouse pointers on shared documents and things that makes us a little bit unique. So there's three traffic types there, right? So you've got your HTTP requests coming in. So we open up a WebSocket for real -time requests and then WebRTC for collaboration. So three main traffic types. And then within each of those, we sort of break it down a little bit further. [5:23] So on HTTP requests, you know, telemetry comes over HTTP requests. Not particularly interesting traffic. Our APIs come in over HTTP requests. Much more interesting to us. Obviously, it has to be... Ah. low latency, high performance if you're doing APIs. Just broadly, in terms of architecture, I think the other thing you didn't touch on is that initial JavaScript bundle that goes back up has the first lot of hydration, and then after that, the APIs are used to hydrate the rest of the space. Yeah, we do the bootstrap, which is like a big, it's an optimization where you send one big payload that starts the editor running, right? optimize the kind of time to interact them on our editor or home page. So we do some kind of aggregation of multiple API requests that we do at standard startup and we do that in one big burst and try to get it out to the user as quickly as possible. [6:16] We're extremely focused on getting the user to using the product as quickly as possible. So down the HTTP request life, every request that comes into Canva transfer. a worker and so that means for us we process trillions of HTTP requests a month inside that worker we run a router and so depending on what we find when we open up the HTTP request so hopefully give us the HTTP request we've got access to the full to the full request we then run a routing layer inside the worker which steers traffic to different parts of our origin so in broad terms you can think about apis coming to our gateway and you can think about assets going to s3 the the routing table is much more complicated than that um it has around 1500 entries i think it's a lot um in a tri structure if anybody remembers yeah cs201 uh cs2 yeah um it's the same in australia cs2 um so if you're gonna try uh and yeah workers, I think the P50 is around 2 milliseconds, between 2 and 3 milliseconds. [7:33] So three different traffic types come into the Cloudflare Edge and then go through this workers-based steering component and then hit your origin. We can keep going down. Yeah, let's keep going. Once it hits the origin, then how does it work? So at the origin, we run a couple of different pools of layer 7 load balancers that route to microservices behind them. Broadly. we've moved to an Envoy -based architecture with our logic kind of sidecar setup. Historically, we had our own gateway that was written in -house, and we moved to open source because it works really well. Behind that, there's a lot. We have a massive microservices architecture. I think the Envoy configuration will route to maybe 250 or 300 different microservices, and then behind those, there are even more things. [8:26] So at the gateway, so we come in from the edge, we come into the gateway, the gateway's envoy now. We look aside for some auth, which we've been doing for a long time. We do some protocol translation inside the network where we're gRPC, outside the network where HTTP. So the gateway does that protocol translation for us, does some work around streaming to convert into gRPC. [8:58] And then on the way back, obviously, send the request for that. So where do you feel like you're working around versus working with Cloudflare in that flow of transactions? And I say this knowing that you have some of the highest scale and a single domain and lots of things that make Canva very special in terms of Cloudflare's customer base. special in a unique way. Everybody's special. I mean, there's a lot. So one odd and, well, I guess interesting thing about Canva is we run just about everything on www.canva.com. And so we have some great benefits from doing that. But also, there's massive problems with cookie payloads. And we do lots of progressive routing. And so tuning things there is an interesting aspect where sometimes we work well together, and other times Cookies are expensive and problematic. [9:53] We also use the WAF to protect us from a variety of things, and it comes with cookie payloads, which is good and bad. So we've been doing optimizations around that. So that's one aspect of it, I think, where there are good and bad aspects of the site. I'm trying to think of other... I mean, there's a set of stories here as... Canva has evolved, we've asked for various things to evolve on the Cloudflare side. So an example of that, we had a worker and all traffic was running through the worker. So I've been at Canva about three years. And when I arrived, the P50 was much higher. We were doing more work in that main worker and the code wasn't as efficient. [10:42] We couldn't re-architect that worker. because of how the workers' billing model was. So this is the interaction between the software side and the Cloudflare platform. And this story's got a good ending, so it's okay for Cloudflare. We spoke to you about that. We now have been able to use workers and sub-requests. We've been able to re-architect our worker and bring that router to the... to be part of the main worker and other parts of the work that we need to do are happening in sub-requests. And what that has allowed us to do is run that router on the edge, but also we can run it offline and process all the HTTP access logs and do route attribution for cost inside the organisation. [11:38] So we are still working around... fact that at our scale it's important to know historic trends of which routes we hit and how much they cost us to run but you know we continue to evolve the platform to help us solve some of these business problems so it was like public facing routes or internal like microservice public facing public facing just so what we've tried to achieve with the programmable edge is is to be able to scale what teams need to do and give them safe config in order to be able to configure their route how they need to. So we've unified our configuration and we've given each team inside Canva, if you want to think of each team owning a route, we've given each team in Canva the ability to configure those routes. [12:34] We then bundle all that up and ship it into the worker. um other challenges we've had with that model uh up until relatively recently workers didn't have progressive deployment so we had to work around that we ran two workers and had to do some interesting tricks for traffic splitting so that we could a b um our two versions of the worker um testing is difficult you know so there's been a bunch of different um challenges you the platform keeps evolving as we evolve so yeah we're happy with that that's good um you know this is sort of a follow-up to that was there a specific spot along that curve where you feel like the relationship with cloudflare changed it was not like you know a component that you had picked but it was more of a strategic partner that we needed to start working more closely with and and thinking about like a longer horizon of working together so an interesting question for me because i arrived three years ago and my sense of things at that point was that canva was uh very much treating cloudflare more like a vendor and less like a partner um our account managers here megan um we've worked hard to make sure that that's not the case yeah um and so we as we start building things we consider things from the cloudflare menu um that menu is getting bigger every day so you know there's 150 things to choose from now um and that [14:07] that menu is quite large but i think i don't want to kind of say when i arrived it changed but it feels um it it feels like it's changed um quite dramatically in in the last kind of two or three years that's great so if we go back to kind of life cycle for question these different workloads are going through there how do you think differently about if at all latency, you know, global distribution of traffic, do those have different, you know, thresholds for you? Or is it all sort of like, it's all one client experience? You think about those streams in a similar way? I mean, I think we're hyper focused on our users and our users experience. So I've been trying very hard to tie our per request metrics into product level metrics and optimize the things partially because the business looks at the product level metrics. [14:58] and they want to make those better. And when I say I made a request, P99, I go down by 5%, they're like, what do the users see, right? So we don't have strict micro -latency kind of targets for things. We have some guidelines, but we mostly look at the user-level metrics. And as we get more sophisticated, we're also looking at user-level metrics with distributions, right? The P50 user experience versus the P99 user experience is vastly different. i've been advocating very strongly for bringing the p99 way down right um that takes more sophistication but but again like we want to make it the best experience for our users and make it as fast as possible um and uh the p50 user has a pretty okay time but the p99 user is not super happy right um in that direction this is a dynamic product that people do creative stuff in it how do you think about caching how do you use our cache with something that's so fluid and dynamic. We try to cash everything we can, right? [15:54] Surprisingly, so I come from a media background before Canva and cash hit rate's hugely important, obviously, for that. Our cash hit rate is actually 70%, which I think for a product, what's 68 % or 69%, for a product that is highly individual, that's a pretty magic number to be able to hit those kind of cash hit rates. One of the good news stories with Cloudflare is we don't have to think too much in the nicest way. We don't get charged for how much we cache, right? And so we do set cache control headers on traffic as appropriate. [16:39] The Cloudflare Edge handles it for us. We don't think a lot about whether we want to or not. We just set the cache control header and hand it over to you guys. in the in the response and off it goes when you start to break down different asset classes or different traffic types you can start to see some really wild differences in how things are cached fonts as an example has something like a 99 % cache hit rate and then there's javascript maybe more like 25 % so you get some some very different numbers and then you have to What is the user actually doing and should we be tuning that up or down? We've actually just done some work the last two weeks or so where it'll actually be net beneficial for Cloudflare because we've changed how our JavaScript is broken down to provide more stable caching of the JavaScript up until a while ago. [17:46] week or two ago every time we did a deploy the the fingerprint of the javascript would change for all of the javascript which on the cloudflare network you'd actually be able to see us when we did a deploy because everything got evicted and got refilled across uh every node uh that won't be or it's not the case now since uh since our changes so yeah um in in that sort of request thinking, what parts of that end-to-end experience are synchronously blocking through Cloudflare? I think you might have covered some of this. And what parts are you able to put off of the hotpath and come back to either post-processing or async [18:31] We have some async API requests and callbacks. A lot of that stuff happens over the WebSocket because we can do whatever we want there. I'm trying to think of, like, what We plumb up chunked uploads and those kind of things. And often, they'll come with a transcode kind of component that comes back. [18:52] There's a lot with the API requests. So actually, amusingly, most of our API traffic is limited to 30 seconds. And so it forces applications into that footprint. And then people push around it in different exciting ways. How do you think about? performance measurement in such a complicated like is it fast enough you said p50 p99 for certain components but like just that end to end when you think about answering that question of what's the user experience that's a hard one to answer it's really difficult yeah yeah um because so canvas mission is to empower the world to design we we take that pretty literally um and so we can see massive differences in performance in uh different locales um and so So sometimes it's the ISP into Cloudflare. [19:43] Sometimes it's Cloudflare's backhaul network. Sometimes it's the interconnect between Cloudflare and the ISP. Quite often it's our code, like we're not perfect by any means. So figuring out those latencies is kind of a full -time job. So how do we think about it? I think at a really high level, we get as much telemetry out of Cloudflare. as we can and we overlay that against our our user metrics um i was running some numbers recently where we spend a lot of our bandwidth on cloudflare on on telemetry of various kinds so we bring all that back into a big data warehouse so we we look at that uh and we try and make it as fast as we can uh from from a network point of view and then we hand off some of that our application engineers and say hey like this bit's not you know it's too fat you need to slim down the amount of javascript or you know provide stable hashes as an example to speed up the application i mean we into we we do the low level bits and bytes but mapping into the product and the usage like we an example is we're trying to compress as much as we can to send fewer bytes on the network right um but i also talk to product people about it's better to send no bytes than to have to compress bytes, right? [21:03] So if we can reduce the payload or reduce duplication, there are lots of things that we can do in the product. And I'm happy to do compression, but there are also product things that we need to push back on. Telemetry is another example where we aggregate a lot of telemetry. We were batching inefficiently. We weren't compressing. There are lots of ways to optimize the code to make the network's job easier, right? And sometimes the problem is that we have to use the code to make the network's job easier, right? people don't understand because the product works okay and i'm like i mean one fewer bite on the network is one fewer opportunity to drop a packet to have problems right so like we it's better for everyone if we send less data right if we have to send it we should but um so we i rove around the stack advocating on behalf of like we can make this so much better like there's so many things to do we have tons of data um and tons of ways of analyzing it uh i try to line up with the standard product metrics that people care about because because that's how the business runs um but then also try to pivot to like look at these latencies look at this distribution uh often people like to make decisions where they're like oh new feature x it like makes everything better but like often it's like well it makes this better but that worse and the decision is gray and so i try to engage with people get metrics that we understand that we believe in true because you can also completely deceive yourself with metrics very easily and just work the entire system. [22:34] Everybody has to collaborate to make this better. But it's a lot of translation, too, between the different layers and getting the people at the right context to do the right work. I know from working with you that Canvas started as a web app, but your push is more focused on mobile end user use cases going forward. How much of that changes when you start thinking about you know thick client and mobile client versus the react based stuff yeah i mean it's an active area for us especially on mobile we've got teams now who are looking at rewriting things that traditionally would have been done in javascript inside the web view closer to the metal if you like we've got people who are looking at redoing various components most of that looks like it's Rust, and then we've got a binding layer with FFI, if anybody's familiar with that, to start binding in some of those high-performance components or higher-performance components into the mobile app. [23:39] So it's an area of active investigation for us, and we're looking at which parts of the mobile stack, as an example, make sense to move in. those lower level languages and still more to come and you know we haven't released an app with that yet but if you think about something like video how much video can you do on on a client and then we start to then think about well that's great if you've got a two thousand dollar iphone but if you've got a hundred dollar android device in [24:19] Nigeria or Philippines or something like that you're probably not going to be offloading work to the GPU and that's where we can start if we write some of these low-level components in a systems language we can start to have conversations about running it okay well it's not on the device maybe on the edge and then if not on the edge maybe in core compute and so we're thinking now about because scale and speed are important we're thinking about portable portable libraries and portable constructs from our edge to our gateway and our core. That's a really good pattern. That makes a lot of sense. And I admire the, it would be easy to focus on the $2,000 phones and it's a much easier problem to solve, right? To actually solve for the world's creators makes you solve the harder part of the curve. That's, yeah, that's our mission. [25:07] And it's some pretty cool engineering when you get involved in it at scale. It's good. So when, speaking of building new things, building a new product you have a new idea how do you how and when does cloudflare come into that and sort of the follow -up is do your development teams know they're using cloudflare or are they using like an abstraction or library or whatever that's sort of like or a little bit of both depends on the use case i'll set up um i'll set up this and then i might head over to las who develops every day um canvara is about 5 000 employees about two and a half thousand of them where some type of technology had. So we're a pretty big development organisation. [25:52] Cloudflare is one of a number of, you know, SaaS vendors that we use, and the way I think about it is they are part of our toolbox, so they are definitely on the paved path. I'll hand over to Laz, but we have an infrastructure... abstraction layer that we've written in -house where we can abstract away some of the complexity of deploying onto various clouds and deploying into various places. It's called Platy. For those who know Canva or know any Canva employees, you can have discussions with them on Platy. Platy's short for platypus, which is an Australian animal. Nobody needs to know that piece. You're all here for the technical piece, but I'll hand over to Laz about. [26:40] what constructs we abstract and and how how users use them internal users i think we have we try to get engineers to build using the most appropriate platform and tool choice is always part of this right um and depending on where we are in the life cycle if it's a crazy new idea that's driven by a product that we don't know will be successful we will try to launch it as quickly as possible uh to get feedback on it right get it to our users um If it's something that we know will be a strategic longer-term vision, then we'll pick different platform trade-offs. An example of a recent note is we launched the Canva MCP on Cloudflare because it allowed us to launch it as fast as possible. Instead of building that technology in-house, we could utilize a Cloudflare abstraction and just go, right? [27:26] It was fantastic, I think, for everyone involved. In other places, if it's a bigger investment or we can't... buy it from a platform, and then we'll in-house it. We're doing the build versus buy conversations constantly. We do have abstractions. And often, many of our product engineers, they know Cloudflare, but they don't know how the caches work or any of that stuff. They don't have to operate at that level. But certainly within the infrastructure, everyone knows about Cloudflare. And then once things take off, you have an idea, you rapid prototype it, you launch quickly, we pay some money for our platform or whatever, and then it- catches on and it's large scale and it's expensive or slow or then we optimize right we double back six months a year later on that product and we change it so that it's efficient to serve it's faster better stronger on the product side they're always innovating on the product stuff but we're also our infrastructure is refactoring itself and moving workloads around the network for the most appropriate place and sometimes that's performance sometimes it's cost sometimes it's availability A big kind of macro theme and infrastructure at Canva is being flexible as we continue to figure out how to do our product better, faster, stronger. [28:39] The world's also changing around us, right? So we're adapting to new AI demands. We have to have GPUs over here. And as that pops up, the engineering organization that serves the product has to adapt to these things, right? And so more and more so today, we're thinking about out not one platform not one product but how can we bundle a bunch of these things together with abstractions so that i can be flexible about where i run my workload um how things find it how we store state how we distribute state um and to do the computation the storage the communication at the right spot in the stack but again early product like launch it figure out how it works and then once it's once it's caught on then we iterate, and refine, and get good at it at a larger scale. [29:32] Often, the product people, I love product people, but they'll come and they'll be like, they're like, Sit right here, man. They're like, this is a great new feature. It's going to launch. The whole world's going to use it. He's in rainbows, and tell me about it. And I love their enthusiasm, and I'm like, I mean, I've seen this a million times. But we enable them to launch in small-scale ways, to get feedback early. do like rapid, get that in front of people and figure out how it works. And then if it's a success, then we enable them to do it at scale, right? But the important part is figuring this out and getting the technology out, and then iterating, iterating, iterating. So speaking of new stuff, then there's a couple of things we've been working on together that are your, as I mentioned earlier, helping push our platform forward, because Canvas scale always helps us do this. [30:20] Can we talk a little bit about some of the two or three projects that we have in flight with you? that partnership's working and sort of like areas where we're collaborating together. Want me to start? Sure. We've spent the best part of the last 12 months re -plumbing a whole bunch of our north -south traffic infrastructure. We touched on some of it just before with Envoy and the place we're going next is actually very much blurring the lines between where Cloudflare stopped. and where Canva starts. So from a network point of view, we're trying to work out a control plane where we'll be able to control the distribution of traffic directly in Cloudflare down into Canva protocols, XDS. [31:10] And we spent a long time, or not a long time, we spent the last weeks, months, working with Cloudflare on that. And it's looking, shaping up really exciting. Laz has been deep in the detail of this, so I'll hand it over to him. I mean, it's the flexibility strategy. So a year and a half ago, I started at Canva, and we were deploying service mesh and trying to figure out how microservices could find other microservices. [31:40] And the routing of that traffic and figuring out how that whole system works turns out to be pretty hard. And as we kind of iterated... it and we're figuring things out we stumbled upon xds and envoy speaks xds grpc speaks xds and so we with our microservices had a strategy for routing and requests we're using envoy as our gateway so we can also route requests there and so we decided we would kind of heavily invest in this as a strategy and so the next step out is we have the edged origin hop that we tell cloudflare we're like hey here's our dns name for our origin and you pull it and we're like well, you could talk XDS, right? And so we've been working together to develop this so that we can use this flexible control plane protocol that in its simplest form is just a replacement for DNS, like here's the IP addresses, but it also comes with here's how to talk to me, here's how to retry, here's how to route this left, right. [32:36] And as we build it up, I think it will be important strategically, but it's a core part of our strategy is as our resources move around the network. we get gpus and a neocloud we need a way to route request to them and or levering these open standard protocols and plugging components into them um and cloudflare is a huge part of this um yes that that's one that's very network heavy i think um higher up the stack um there's an internal tool that we're um developing at the moment uh for for canvas staff based on real time um as well So we're not 100 % sure whether that will be a product that goes into the main part of Canva. [33:21] But 5 ,000 staff is a relatively large footprint and we've got a bunch of software developers. So we're looking at moving some things onto some tools that we're developing in collaboration with Cloudflare. I can't talk too much about it because it's not necessarily... going to make the public Canva product, but it's an example higher up the stack as well where we're taking primitives from Cloudflare and, you know, stitching them together in some interesting ways. So there's many stories like that, but that's one that's higher up the application stack. I know, like I said, this is exciting for us because, you know, the scale that you operate at, the regions, the ASNs, the networks, the devices, every aspect of Canva. [34:09] pushes us. Even the fact that it's all one domain pushes us to think about our product differently. So I really appreciate it. I got one more question, then we have just a few minutes for audience questions. If you have something in mind, get excited about it. And we'll get a microphone to you in a minute here. If you could get one thing out of the Cloudflare team, I mean, you have a straight line to a lot of people, so this is not really hypothetical. But if you get one more thing out of the Cloudflare team, what would it be? [34:36] Well, I I've got my wish already. You can think. I'll think. So I've been doing a bunch of, I'm like a protocols guy, right? I've been doing a bunch of work trying to optimize things. And so the one wish that I would have today would be to tune the H-Pack buffer size up. HTV2 has this compression thing for headers. You get 4K that's shared between the client and the server. We have 6K of cookies. And so that. 4k is useless we just thrash that cache on every request and so we're sending an extra 6k essentially with every request to our massive single domain i can't get them to delete cookies on so if we could get it to be 8k or 16k and i could fit it in there then every request would be smaller and so that would be my one wish but then i'm also rolling out http3 right now and that one has the the shared header uh Dictionary, it defaults to zero and Cloudflare sets it to zero. [35:38] And so I would like it to be on zero for similar reasons, right? Like it's about performance and optimization, right? [35:49] We have some capacity constraints limited by the physical world. There are places in the world where it's really, really hard to get a good bandwidth. And there's a chicken and an egg problem. because good bandwidth can often lead to GDP uplifts and a whole bunch of other things in kind of some countries. I would love Cloudflare to build out more capacity in some of the places where we know we have problems. And I don't want to point to any specific markets or regions, but ahead of us being able to expand into those regions. So when we talk about empowering the world to design, we mean everybody. And if you can't get a good connection for a cheap price, that really inhibits your ability to use the product. [36:37] So we're working as fast as we can to reduce the weight of the product. On the capacity side, there's some places that would be great to have a bit more capacity. Both very high level and low impact and network infrastructure. Okay, we have a few minutes for questions. Anybody have a question in the audience before we wrap it up? We've got one in the front here, please. We've got a microphone coming, so just wait one second. Thank you. At such a large scale, have you guys been able to utilize Cloudflare's durable objects when you have multiple users on this same application You want to talk about it? I kind of don't. You can talk about it. We are not using durable objects. [37:23] We have been able to do this. a deployment in china there's canva .cn and durable objects are not available in china so we can't lever it there's some there's some small use cases where we do use um do for for a couple of events and things um the other part about camber just to know for the room um we own a set of companies so camber .com is by far the largest amount of traffic but um we own a company called pexels a company called pixabay a couple other companies affinity and some of those have some small deployments of durable objects but on the main product we're not a big durable objects consumer we've played with it yeah we know what it does yeah it is an example though of how you push our platform that we have a china network product that lets us work your partner and deploy you know canva.cn for a customer like this the menu of cloudflare does not all automatically work in that secondary china-based deployment and so It's another vector that we work with y'all on and say, yes, it works in China, but we need this and this and this. [38:27] It's a complexity for us to manage. It's a tricky situation. And we try to keep it as homogenous as possible, but it's difficult. One last question. Alex Krivet. [38:40] Hi. My question was, you had mentioned both, you don't want to design things for just the users with the $2,000. devices at the same time canva has such a wide breadth sometimes you're operating in really lossy networks across the world how as a company does how does does canva approach that from an engineering perspective how do you build empathy inside your engineering org to make sure that you're you're building for those people who are probably in well connected in well connected devices when they're building but maybe deploying to less well -connected places um i've got a recent example of this um you can actually go inside our resource booking tool and book a low -end device so we've actually got a little in in sydney there's a room um that's been set up with a bunch of relatively inexpensive or low powered uh low -powered android devices um the the wi -fi access point in that particular place uh also has some uh some clipping done to reduce the bandwidth um and we got contacted i didn't do my work and actually gave it to las um we actually pulled down some stats from the cloudflare network and gave them to our it um our it team who provisioned up like a p50 uh bandwidth um so that when you go and pick up that low power device you're getting close to um the [40:14] the performance of what that person might see. That's an example of where, as an engineer, if you're working on a particular feature, you can go and grab that device and play with it. When we present things internally, often they will show a video of a lower-end device. It's a high focus for us right now. The other thing that I like to do, because I'm a network bits and bytes kind of guy, is I explain to the product people that every byte we send costs money, and they're like, I mean, do you want to cost... reduce our bandwidth and I'm like yes but also some of our users pay per byte right and every byte you charge them like is is an aggregate bad right so if we could all save money if we send fewer bytes um and I guess people in the U .S. I have a flat rate plan it's unlimited etc etc so you don't really think about this stuff but in other places bandwidth is very expensive and if we're wasteful with it it costs people legitimate lots of money right and and uh I try right like you you have to understand this is this is non-zero amount of money for people we should be efficient um and we're when we're building the application we're starting to do things where you kind of adapt with the device or technology so we'll build the you know high pollutant thing for the iphone and then we will scale it way down for android to be very efficient right um try to give similar product experiences but we didn't actually do this but i did threaten at one point in our non-production environment and because we run so much traffic through the worker i did threaten just to put in a sleep yeah and and basically if people were testing in [41:48] in non -prod um and then and then tune that number so like this week we're operating as if we're in the philippines and next week we're operating as if we're in nigeria so those toggles are available to us but um at the moment the way that empathy is done is through education and a little bit of device testing. Plug for Ryan's Beacon product over here. We just announced the birthday week. [42:16] It's web metrics that are not tied to Chrome that are available for your data warehouse type lookup because they're in a giant BigQuery table. So you can actually get benchmarks and baselines anonymized and have your product compared to those things. So if it's slow in this entire country or you're just slow in this country, and that's the sort of thing that this public data set... us at birthday week one of the interesting hallway conversations we've had since we've been here is take everything that you know about uh rum you know and measuring and then go and what about the agents how are we going to think about that same uh accessibility question around yep [42:55] um but but the the point is the same which is you want to provide a good experience so start with I'm trying to create the empathy and try and measure what you can nice we did make it almost all the way through the whole talk without saying agents I'm glad you got that in it's mandatory I think it's mandatory it's mandatory we can't do a talk of this we talked about MCP though so okay indirectly directly well that's all time we have thanks everybody for being here really appreciate it thanks Michael and Laz for participating ## Related sessions - [Co-engineering the future: How Uber & Cloudflare engineer a global platform](https://www.cloudflare.com/connect/2026/session/co-engineering-the-future-how-uber-and-cloudflare-engineer-a-global-platform) - [Scaling Discord’s voice on Cloudflare’s edge](https://www.cloudflare.com/connect/2026/session/scaling-discords-voice-on-cloudflares-edge) - [A runtime for developers and their agents](https://www.cloudflare.com/connect/2026/session/a-runtime-for-developers-and-their-agents) ## More from Connect - [All sessions](https://www.cloudflare.com/connect/) - [Get notified for Connect 2027](https://www.cloudflare.com/connect#2027)