John Graham-Cumming, CTO of Cloudflare, joins Ben and Ryan for a conversation about the latest trends in internet usage highlighted in Cloudflare's 2024 Year in Review report. They explore the surprising growth of internet traffic, the implications of post-quantum encryption, and how mobile traffic differs between iOS and Android. The conversation also delves into the rapid adoption of Starlink, the rise of AI traffic, and the future of internet growth.
Check out Cloudflare’s 2024 Year in Review.
Read John’s posts on the Cloudflare blog or connect with him on LinkedIn.
Shoutout to user Timo Kähkönen for providing knowledge-seekers with a cheap algorithm to find measure of angle between vectors.
[intro music plays]
Ben Popper Hello, everybody. Welcome back to the Stack Overflow Podcast, a place to talk all things software and technology. I’m your host, Ben Popper, joined, as I often am, by my colleague and collaborator, Ryan Donovan. Ryan, I feel like you cleaned up your background.
Ryan Donovan Uh, maybe, maybe. It’s just–
BP Things are looking ship-shape there.
RD A new, new picture in the background.
BP Yeah. Well there's a certain symmetry to it, at least.
RD Mm-hmm.
BP It looks ordered. We are, uh, very lucky today to have a guest on, John Graham-Cumming, who is the CTO over Cloudflare, and we are going to be chatting about some of the big picture trends that they observed in their 2024 kind of state-of-the-internet report. Ryan, I think, uh, you know, there's certain things here that jump out to me as intuitive and others that are not.
RD Mm-hmm.
BP I didn't know the internet was so seasonal. Did you know that the internet was seasonal?
RD I, I didn't know it was still so seasonal. I remember, you know, back in the day, it used to be the, uh, the September crush, as all the new students came on.
BP Ah, I just feel like everybody's always online. I don't get how the internet could be seasonal anymore, but we'll learn about that. And then the other thing I– really piqued my interest. I know, Ryan, you and I have done this, is the post quantum encryption age is coming.
RD Yes.
BP Uh, we gotta get, HTTPS is not enough anymore. Anyway, without further ado, I would like to welcome John onto the podcast. John, welcome to the Stack Overflow Podcast.
John Graham-Cumming Well, thank you very much for having me.
BP So, for folks who are listening, just uh, give them a quick flyover. How did you get into the world of software and technology? And what led you to the position you're at today?
JGC Ha. Wow, that's an interesting question because when I got into software and technology in the UK in the eighties, it wasn't obvious, at least to me at school, that computers were going to be anything significant, you know, and of course, I was wrong about that. So I actually did it just because I loved– I was a, I was a child of the 8-bit era, and so I ended up playing around with 8-bit machines, and that's what got me excited. And I'm, I'm also a child of a thing called the, uh, Computer Literacy Project in the UK, which was the UK's attempt to teach us all about computers. And I had a thing called the BBC Micro, which was a computer made by a company called Acorn, which became Arm. Um, so I was part of a, a government effort to, uh, teach us all about computers, but I think I probably would have done it without government help because I was fascinated.
BP Yeah, I, I love that history. I also recently heard sort of, maybe, the next turn of the wheel there, about the UK Biobank, which is this incredible longitudinal health study they've been doing, and now, is far and away the most popular sort of a digital data source for people doing biomedical studies around the world. So the UK has invested in some really interesting things that have paid off over time. Tell people who are listening, many of our listeners are my age or younger, what's it like to live in the 8-bit era? How do you make a program useful and fit it, uh, onto, not e– not even a floppy, I'm a floppy era. I’m 10 years after you.
JGC I– yeah, I mean, I, I have, you know, you're talking about kilobytes of memory, right?
BP Mm-hmm.
JGC So you're talking about having 16, or 32K, or 48K, if you were really a big spender. Um, obviously amounts of memory that are smaller than, uh, you know, a JPEG, uh, that you pass around with social media. I mean, I don't think it was– in a way, it wasn't fundamentally different. Uh, it's still programming. Um, you, you always have constraints. Those constraints might be how fast your processor is, uh, how much data you can handle. I think the good thing was that the entire computer was knowable, basically.
RD Mm-hmm.
JGC You know, I, I mean, I grew up in an era where my BBC Micro actually came with the circuit diagram, um, and, you know, you could quite easily disassemble the entire operating system and understand it. And all of us did. And there were books about all of the details about how the thing worked. That's not really possible anymore. I mean, understanding how this Mac that's in front of me actually works in deep, deep detail is, I think, beyond the single human, uh, to be able to have that context in their head.
RD Yeah. I, uh, I’m also, uh, of the 8-bit era. I started on a Commodore 64, so remember the days of, of single person, single, individual, people doing programs.
JGC Yeah, absolutely. I mean, one of the things that was interesting about that era and– was you got individual programmers who became, you know, wealthy off of one game, or two games they created, or something like that. It was, uh, I was utterly fascinated by that, at the time.
BP Yeah that, uh, movie that came out recently, the Tetris movie, uh, is really amazing. The story of how that was created and passed around and, uh, just sort of how one individual's brainchild. So, to sort of change things up a little bit, we're going to be chatting a little bit about the Cloudflare, uh, you know, research that comes out once a year, which looks at internet activity. And now, we've gone beyond, I don't know, exabytes, I mean, we're, we're at the, the largest amounts of data flowing that, that one could imagine. Can you tell us a little bit about how this study is conducted, how you gather the data, and, and what you– you know, what you and the team there do to sort of pull insights from it?
JGC Well, of course, Cloudflare operates a huge number of websites around the world. That is to say we do, uh, security for them. We do acceleration for them. We think it's something like around one in five websites use our services. Uh, plus there's a bunch of other stuff we do, uh, including our free DNS Resolver 1.1.1.1. That causes us to have a lot of data about what’s happening, especially in terms of trends around the internet, so we can look at the growth of traffic, we can look at the popularity of different services by looking at the DNS activity. We can get a sense for, you know, what’s happening with different protocols. You mentioned post quantum before – that's a new, really interesting area. Um, so it's really from our own data because of the use of our service by our customers, and we publish a lot of that data. Now, it's all completely anonymized, of course. We're not looking at what individuals are doing. We're looking at overall trends. We publish it through a thing called Cloudflare Radar on our website – radar.cloudflare.com – and people can go in, and they can, you know, figure out what's happening in a particular country. Like, for example, one of the things we published not long ago was internet quality information, so we can see how fast people's internet connections are. So you can say, “Hey, I want to compare France and Germany or something,” and you can go in and say, you know, who's got lower latency, who's got higher bandwidth, or whatever in their ISP, you know, availability in those countries. So we have a lot of data of that style – tends to be, um, high level trends, but within those trends, there's lots of fascinating things to find.
RD Yeah, and I think, before Ben and I sort– you know, ask about the things that we found interesting. I wonder if there was anything that you saw in there that, that surprised you or stuck out to you?
JGC I don't know if there were things that really surprised me. I mean, there are things that are perhaps surprising when you, when you stop and think about them for a moment. I mean, we saw the internet traffic grow 17% this year, and it's been doing that, something like 20% per year for years. Which is incredible when you think about the compound effect of that. I mean, the internet use is growing enormously. And then if you look at things like Starlink – because we can see the traffic coming from Starlink to our network. Starlink is growing very rapidly as well, and tends to, uh, you know, tend to take off in countries when it becomes available very rapidly, uh, and, overall, is growing very quickly. So, that is perhaps, is perhaps, if you have never looked at this stuff, is a bit of a surprise. Um.
RD Yeah.
JGC One thing that's a surprise is just how much bad traffic there is on the internet. That is to say, attack traffic, bots doing stuff we don't want them to do.
BP Right.
JGC It's, it's just constant.
RD Yeah.
JGC And that, that always blows me away, you know.
BP It's my washing machine three times a day, seeing if it can, uh, disrupt my neighbor. Blah, blah, blah.
JGC Yeah. My washing machine talks to the internet every day about something. I don’t know. I don't even use it every day.
BP Chatty. These are chatty machines. Yeah.
RD Mm-hmm.
JGC What are they talking about? So.
BP Yeah, I mean, I guess I'd be curious if you have an opinion on this. You know, you're saying it's amazing how much the internet continues to grow, obviously, right? The, the population using the internet is growing, but 20% year over year, 17%, there's a– some of that traffic is, you know, just consumers. Some of that traffic is enterprise. Maybe some of that traffic is, uh, AI crawlers. I know you mentioned those, and, you know, uh, what you're looking at in some of that traffic could be, you know, the bots and, and the cyber. Do you have a sense of what's driving the growth, like, what the principal driver of the growth is among those factors?
JGC Oh, I mean. Uh, uh, uh, it’s, first of all, it’s us, right?
BP Yeah.
JGC It’s not, it’s not bots and things like that. Obviously, the bots are an important part of the traffic, but they’re not the thing that’s causing overall traffic to increase. You know, I mean, we're slapping the internet on everything, right? As you, you know, as you say, your washing machine is talking to the internet. I can’t– I have countless things in my house that would like to chat on the internet. Plus, there’s our own use of it for absolutely everything, particularly since Covid. Uh, you know, people have moved so much stuff online, and you know, if you look at statistics about people working from home, the amount of, like, traffic that's coming from home networks because people are working at home.
BP The three of us are doing it right now. Look, look at what we're doing.
JGC Yeah.
BP I used to record this podcast at an office, in person with people, and then upload it to the internet later.
JGC Yup. And I'm sure that my house right now, you know, the washing machine is probably, you know, asking if it needs an update. And I have, you know, a plethora of devices, and we've all got these things. So, my car is probably chatting with the internet about something. I mean, so, we are adding internet to a lot of things, and we, ourselves, are just using the internet for pretty much everything.
RD Yeah. One thing I was sort of scanning over and thought was interesting is: we're sort of mostly using the, the older protocols, right? I think there's, like, 28% of traffic that can use IPv6 is using IPv6. HTTP/3 is only used by about 20%. Why do you think there's kind of this, you know, slow adoption of new protocols?
JGC Well, I think those things are two different stories. I mean, the thing about IPv6 is that it has been around for a long time.
RD Mm-hmm.
JGC Right? You know, we're at, we're at something like 30%, let's say. Traffic is, is happening over IPv6, and, you know, what's really driving IPv6 is mobile telephones, and there are notable mobile phone companies who use IPv6 on the phones, and we see that as the driver. It's funny because I, in 1999, wrote an article for the Guardian newspaper in the UK about how IPv4 address space was going to be exhausted soon. 25 years ago.
RD Right.
JGC Now, the truth is we found more and more creative ways to use IPv4 addresses.
RD Mm-hmm.
JGC And there's a lot of infrastructure that uses IPv4, and so, it’s kind of got stuck, in a way.
RD Mm-hmm.
JGC And the HTTP protocols, I think, it’s very different. Those tend to have a much faster, um, you know, adoption curve, and that's because the devices, the, you know, your browser, your phone, your laptop, they can upgrade the underlying libraries that use HTTP much more rapidly than a switch from IPv4 to IPv6.
RD Mm-hmm.
JGC And so, although you only see, you were mentioning that the IP– HTTP/3 is not, perhaps, as, you know, as high as we'd like it to be. It's only, like, 20% right now.
RD Mm-hmm.
JGC But HTTP/2 is at 50%, right? And HTTP/1.1, which is the old version – also from 1999, funnily enough – um, is still around. And now, what's interesting is, if you look at HTTP/1.1, you often see really old bots using that protocol. So, actually, over time, the use of, you know, an old protocol becomes a bit of a signal to say, “I wonder if this is genuine or not.” You know, if this is a real, real thing.
BP These are the zombie bots floating around helplessly. Not really, not really doing much anymore.
JGC Yeah. Because what happens is hackers copy and paste each other’s code. So they really use stuff over and over again.
BP I, uh, wanna piggyback on Ryan's question and touch on something we've mentioned a few times without diving into it, uh, too deeply, which is the quantum encryption. This was one of the statistics that you pulled out to highlight. What is post quantum encryption? And why do you think people are, are switching to this?
JGC So, let's talk about quantum computers. First of all, we are all using classical computers right now. We’re talking on one right now. And there is a sort of computer which is semi-theoretical, but getting more and more real every day. I mean, you hear announcements from IBM, from Google, notably, about quantum computers they've built. These computers are able to solve problems that take a long time on classical computers, uh, very quickly. And, notably, one of the things they can solve is certain sorts of mathematics that underlie cryptography. So, once you have a viable quantum computer, there are cryptographic algorithms – notably RSA – that become weak in the face of those computers. And so, the post quantum world is a bit weird because we have a deadline. At some point someone's going to have a workable quantum computer and can break cryptography that we use today. Uh, but we don’t know what the deadline is. So, it's a bit like having the Y2K problem, but we don’t know when the year 2000 is. It could be in five years, and it could be in 30. What's happened is the cryptography community has got together and said, “What we need to do is put in place a new algorithm that is safe against these quantum computers.” And that's what post quantum encryption means.
BP Mm-hmm.
JGC Um, and we've been involved in this for quite a long time, actually. Uh, something like, you know, I think it's getting on for eight years, seven, eight years now. Um, helping work with others such as Google on quantum– post quantum cryptographic algorithms. Implementing them, testing them in the real world. And NIST, in the US, has standardized some of these algorithms. And there's a graph on our, on our website showing that there's a big, sudden uplift in, uh, the use of post quantum cryptography in about April of last year. And that was because Google Chrome enabled it by default. Uh, and because Cloudflare makes it available for all of our customers included in their plan. Suddenly the connection between your browser and anything on Cloudflare was using post quantum crypto, and, therefore, was safe against a future, uh, quantum computer. And one thing is you might say, “Well, why am I worried? Why am I trying to be safe from the future?” And the answer is, well, if there's an adversary who has a lot of money, they could start recording internet traffic today, saving on a whole bunch of disks, and just waiting around for a quantum computer to come along and start decrypting stuff. So it's worth doing it now. And what we've seen is actually the sort of forward-thinking security companies, so, like, things like banks, which are worried about these things, are now actually implementing post quantum, and it's a real driver.
RD Interesting. Yeah. I wanted to dig into kind of an interesting, uh, sliced statistic you have there that says nearly one third of mobile traffic was from Apple iOS, but Android has greater than 90% share in 29 countries’ regions. What is specific about those, those regions that Android has such, such a market share in it?
JGC Well, I mean, you see this, right? This is one of the interesting things you see is you see different adoption, you know, timelines and different, you know, centers of usage of, uh, different technologies around the world. As you say, Android will be more popular in certain areas than others. And actually one of the things you can do in Radar is you can dig into those areas and see where those things are. And, of course, now, in China, you've got Huawei having their own operating system. It depends on the, on, on the region. I mean, I think it's roughly connected to the wealth of those countries.
RD Mm-hmm.
JGC Uh, but only roughly. Depends where, you know, Android took off. Like, for example, Samsung, incredibly, uh, you know, incredibly popular.
RD Mm-hmm.
JGC Um, you know, Korean company, and I think if you go to Korea you would see that Android is incredibly popular there. The interesting thing is that, as you say, the peak is really, really interesting.
RD Mm-hmm.
JGC That it's, like, Android seems to peak at, like, 95%, um, and iOS, lower. And the question is: what are people doing on their Android devices that they're not doing on their iOS devices? And I think the answer to this is that Android gets used for a lot of things beyond personal phones and tablets. So you'll often find that, you know, your in-car entertainment system is Android under the hood.
RD Mm-hmm .
JGC Or, you know, something in your hotel room is Android under the hood. So, I think, overall, that explains why you see peak traffic be different because, in fact, it's not necessarily an Android phone. It's uh, it's some embedded device.
BP Yeah, I mean, thinking about how the internet is evolving and how adoption evolving is fascinating. The statistic you mentioned earlier on, which I'd love to dive a little deeper on, is the Starlink one. Ryan and I have gotten the chance to chat with a few folks from SpaceX and Starlink over the years. I, myself, live in a pretty rural area, and my dream is to have my fishing hole and my, uh, remote office be the same place. So, you know, Starlink, uh, is appealing to me from that perspective. And I was spending some time looking at their various offerings. They've gotten incredibly small. I mean, like, it's, like, you know, something I slip in my backpack, can use for most of the day, and I get a hundred megabits up and down. So, no podcasting, but everything else seems great. What does the adoption curve look like in these, uh, regions that you've noticed? And what's your perspective on where things are headed in the near future?
JGC Well, I mean, one of the things that's striking about Starlink is the worldwide increase in Starlink traffic is 3x.
BP Mm-hmm.
JGC In one year. So it is growing very, very rapidly. And then, if you go into an individual country, you know, where Starlink becomes available, you see this incredible shoot up. So you can see that there really is exactly your use case, which is there are people who are desperate to have, you know, a really good connectivity wherever they are. And so, you know, Starlink is pretty successful, uh, in terms of the amount of traffic that's being generated. And we know that, from our own measurements, the latency is very good, so people are gaming over it. And, you know, if you've got 100 megabits, I think you could podcast over 100 megabits. That's, uh, that's quite a bit of traffic.
BP Yeah. Maybe you're right. I mean, right, I guess what it's being used for is, is definitely an interesting question. Does, does your data get that granular? I know you anonymize it, you know, right, like, when you're sharing it, but you're saying, like, you see what people are doing with it, not just, uh, how much they're using it.
JGC Um, if people use 1.1.1.1, our DNS resolver, then, in aggregate, we can see trends around the domain names, right?
BP Mm-hmm.
JGC So we can see like, you know, is Google more popular than Facebook, for example. Um, what we can't do is dig down very, very deeply. Um, we do have some data for different networks, so we can get some sense for what people are using, uh, Starlink for. I don't believe that we have ever actually done that analysis, um, to try and to try and sort of understand all this particular network. Um, and sometimes if the, uh, user volume, uh, is not that great, we deliberately don't do any analysis so that we, you know, there's no risk that we, uh, you know, end up accidentally, um, de-anonymizing someone in a sense. So, like, some, somebody's, you know, usage, uh, becomes public. So, we, we're very careful to do things only when there's a very large aggregate, uh, amount of data available.
RD Mm-hmm. How much of the, the traffic out there is, is human to server kind of connections, like, other than, you know, web crawlers or devices talking to other places, how much of it is actually a human involved?
JGC Yeah. So we can actually find out information about that. So we can look at the traffic data and look at what bots look like. So, if we look at, like, for example, looking live data for the last seven days, we think about 30% of traffic is non-human.
RD Mm-hmm.
BP That is maybe a nice segue to talking about the new kind of traffic that's evolved over the last couple of years, which is AI traffic and AI crawlers. What are some of the interesting trends you've seen here? You mentioned, uh, you know, in the data that you saw a big surge in crawling activity, and then some, uh, of the companies that were doing a lot, like ByteDance and Anthropic fell off while Google continued. What, uh, what do you make of the AI crawling activity? Um, and then yeah, we can dive a bit more into how you're measuring usage of some of the, you know, these gen AI services that, at least anecdotally, seem to have taken some volume away from search.
JGC Yeah. So in terms of the crawler stuff, when we look at this very carefully, because, um, you've seen some crawlers be quite aggressive about looking through websites, looking for new content, looking for content they can build models with. Um, and you’re absolutely right that ByteDance’s Bytespider was very active at the beginning of 2024, and then, over time, slowed down. Now, what does this mean? Well, maybe they consumed everything they needed to consume to build their model. You know, I wonder sometimes if these things aren't a bit cyclical. So, you know, there'll be another round where people need fresh data, and then, perhaps they're building a new model. And you can see, you know, certain services become proper. I mean, ClaudeBot, they got going in May. OpenAI seems to crawl fairly often, uh, I mean fairly consistently – certainly have been since, uh, the beginning, the middle of the year. And so you see different, you know, the different bot patterns there. In terms of usage, it's like anything else, right? It's the same DNS data, so we, you know.
BP Mm-hmm
JGC We can get data, which would let us see, you know, what's happening in terms of, you know, applications people are using. So, you know, are they using Google? And are they using TikTok? And all these kind of things. So, yeah, I mean the generative AI services, I mean, we can, we can kind of do a bit of a compare using DNS data, and clearly, right now, OpenAI is the winner in terms of traffic.
RD Yeah, I, and I thought a interesting bit of that DNS data is flagging certain, uh, top level domains as almost entirely spam. Is it the .bar, .rest, and .uno? Like, I, I don't think I'd heard about these, but is, is that gonna be a, a signal for the future? Just, like, what your top level domain is?
JGC Oh, I mean, that's long been a signal, right? For, so what you're talking about there is our email security service–
RD Mm-hmm.
JGC –where we handle email for customers, then we do spam filtering and other sorts of security filtering for them. And, yeah, and obviously we're looking at domain names used within, say, spam or, or, uh, phishing emails. And then you can look at, you know, how new the domain name is, uh, what the registrar is, and also, and also what the top level domain is. Um, and you certainly do see, I mean, at one point, I think .xyz was very popular.
RD Mm-hmm.
JGC Uh, as, as a, you know, as a, as a TLD. And now, as you say, .bar–
RD Yeah.
JGC –was, uh, you know, was, you know, I had the biggest share of malicious and spam email. Uh, these things move around. But yeah, obviously we would use something like that as part of decision making about whether we think something is suspicious.
BP Just to touch one more time on the, you know, AI stuff, was that at a scale that was significant in terms of, right, a change in what human beings are doing with the web? I know you mentioned, sure, it's grown 17% overall.
JGC Yeah.
BP You know, some of that is my, my washing machine and my car. Some of that is actual people. Um, did gen AI take up a meaningful amount of web traffic for what human beings are doing? Or is it still, you know, a very, very small fraction compared to our classics – social, and search, and things of that nature?
JGC Well, I mean, one way to look at that is to look at domain rankings to see what's sort of in the top 10 or whatever over time.
BP Mm-hmm.
JGC And there's none of, none of the gen AI companies are in the top 10, uh, in 2024. It's Google, and Facebook, and Apple, and TikTok, and et cetera.
BP Right.
JGC Um, so obviously, we're still using all those things, uh, out there, but you do see, you know, gen AI companies getting, getting a lot of traffic and appearing in, uh, you know, sort of the top 100s and things like that.
BP Okay.
JGC Whether it's taken it away from search or not, I, I, I can't tell you. Whether it's additive – that's a good question. We haven't done that analysis.
BP Yeah.
JGC What, what we certainly know is that, you know, gen AI has become something that we are seeing across the board and across the world, in terms of use.
BP Yeah. I mean, I think to say within the first, you know, two years of consumer adoption, you see a few in the top 100. That's certainly indicative of a trend, and we'll see where it goes from there.
JGC Yeah.
RD Mm-hmm. You said, uh, earlier that this sort of 20% growth year over year has been the norm for a while, right?
JGC Yeah.
RD Do you see that continuing? Like, are we, are we maxing out? Is it just going to be a lot more devices and bots doing traffic?
JGC I mean, are you maxing out? ?
BP I'm trying to go backwards. I'm trying to do less. I'll do my– but I, it's hard.
JGC But your devices might not be.
BP Yeah.
JGC Your devices might be, you know, spending all their time, so.
BP Unfortunately, Ryan and I are having children, so we're contributing to the problem.
JGC Yeah, so that's going to add, you know, you're going to add extra users and you're going to add, you know, all sorts of interesting traffic. Uh, but no doubt, you’re going be up late at night with children and looking at your phones, so you're going change the diurnal, so, yeah.
BP Yeah.
JGC You're not helping. Sorry, ,
BP Sorry. We're part of the problem. I guess that though, yeah, the one question I wanted to ask was sort of, like, we've been talking about the web and what you're seeing. You know, you've mentioned some of the things that Cloudflare does. You're the CTO there. Do you want to touch on just one or two of the internal projects that you've been working on over the last year – or that you're working on next year – that you think are interesting?
JGC Yeah. I mean, uh, the thing that’s interesting at Cloudflare, particularly from my team, because my team is both the radar team, which we've been discussing today, and the research team, uh, who work with universities and look at long term stuff. I mean, they, they're the group who started sort of, seven or eight years ago, looking at the post quantum stuff. And we're sort of thinking now about what does the, what is the future of what that team does? And one really interesting area is around efficient use of AI. So, you know, there's been a lot of talk around the electrical cost of, you know, running these AI models–
BP Yeah, yeah.
JGC –doing the training. But, for a company like Cloudflare, because we have always made a success of being efficient in our use in CPUs, right? We now want to do the same thing for GPUs. So that's going to be a research area, like, how do you keep these things running at 80% utilization, because that will save a lot of electricity and, ultimately, will save money for the company.
BP Right.
JGC And so that's a really interesting area, um, to look at what that looks like, and it will have a lot of different aspects to it. Um, some of them hardware, and some of them software, and some of them probably mathematical.
BP Yeah.
JGC So that’s kind of a, kind of a fun area. I think the other area is, as we've got larger and our network has got larger – we, you know, we're in 300 and whatever cities worldwide – the, the complexity of running such a large network goes up quite a lot and building software that allows us to run it has put us in a position where other people can run their stuff on us. Um, so we have this thing called Cloudflare Workers, which has been around for about eight years now, I think, um, which allows you to write code, and put it on Cloudflare, and also run AI models as well. A part of which is that. And so, all of the interesting thing there is, like, how do you take what we know how to do and give it to our customers?
BP Mmm.
LGC And we've done a lot of work on that so that you can scale something without worrying about the scaling aspects of it.
BP Right.
JGC So there's a bunch of areas there that I think are, think are going to be interesting, and we'll obviously will end up resulting in products, uh, improved stability, you know, performance, and also reliability of the system.
RD Yeah, it was interesting. A few years ago, I did a article on the infrastructure behind edge functions and talked to a bunch of folks, and you know, Cloudflare runs a few of those, those programs. Deno Deploy is one of them. I sort of wonder how much infrastructure, how many physical machines, how many regions of the world does it take to actually support the Cloudflare Worker program?
JGC It was probably hard to get that number from anybody.
BP Yeah, yeah.
JGC We tend to be quite careful about that. I mean, because it’s, like, it's all about our efficiency and, you know, using as little electricity as possible. But, yes, I mean it, the other thing that's interesting about the, the kind of edge functions, if we call it that world, is that you are breaking program into very, very small pieces–
RD Right.
JGC –which allows you to actually schedule stuff on machines very, very efficiently and use them, you know, at very high utilization, and charge people only for what they’re actually, you know, using. So actually, I think it's the right thing to do, is to break stuff up really small. Then your infrastructure does not have to be as large as you might imagine, uh, if you were writing these very large, monolithic programs, which are just constantly running – often constantly running – just polling something, saying, you know, “Has anything happened? Do I need to do something right?”
BP Right.
RD Right.
BP Yeah. John, I, I think I've had a few conversations recently with friends who are in the climate space, and what they're interested and hopeful about is maybe something that you touched on: the idea of distributed compute. If people are going to be building out these super clusters that they're going to use for training runs, what can we do with them when they're not in the middle of a training run? Or when they're at 50%? And can that actually be a way, uh, you know, for, uh, us to be more energy and compute efficient at a, at a global scale? So.
JGC Yeah. I mean, one of these collaborators, of course, is we, we have all these data centers around the world where we run everything, so, and we can, we can move stuff around, right, to those different data centers, so we keep them all humming over so they're, it's efficient, being used at high, high levels of, uh, utilization. Very efficient. The other thing we can do is we can, you know, if we are buying green power in some location, um – we actually have this, it’s called green compute. You can go in, and you can say, “Only run my code on green electricity.” And we’ll then figure out where to put it in the world.
BP Right.
JGC Um, so there's lot interesting options there once you, once you break things down to small pieces, and once you have a large network.
BP Yeah. I think that's fascinating. I've seen some people starting to do that within their own home, as backup batteries become more available. You know, send power to the battery when the, it's cheap, or you know when the sun is shining. And then, when it's not, you know, draw from it. And so it's interesting to think about that as, at a global sca– as the sun goes around and the wind picks up, you know, we're chasing the power around the, around the globe to try to stay green.
JGC Completely. Completely. For example, for us, the sun goes down, we– you touched on this at the beginning. There are, you know, trends in the internet, and one of the trends is people do go to sleep. And so, in a particular location, as the sun goes down, after a while, the utilization of the internet goes down. Well, that means servers are being used less. Also means power tends to be available. And so, if you can use those servers during the night, so, like, for us, we might move a workload around and do video processing somewhere. And that becomes more efficient and certainly more environmentally friendly.
[music plays]
BP Very cool. Alright, everybody, it is that time of the show. We want to shout out a user who came on and shared a little bit of knowledge or curiosity. Awarded two days ago to Timo. Looking to find a cheap algorithm to measure an angle between vectors. If you're trying to find cheap algo to measure an angle between vectors, Timo has an answer for you. In fact, the answer was so good, it got more upvotes than the accepted answer and earned Timo the Populist badge. So thanks, Timo, you've helped, uh, over 20,000 people with that one. As always, I am Ben Popper, host of the Stack Overflow Podcast. You can find me on X @benpopper. If you’ve got a question or suggestion for the shows, shoot it on over to podcast@stackoverflow. And if you like what you heard today, do us a favor, leave us a rating and a review.
RD I'm Ryan Donovan. I edit the blog here at Stack Overflow. You can find the blog at stackoverflow.blog, and if you want to reach out to me, you can find me on LinkedIn.
JGC And I'm John Graham-Cumming, Cloudflare CTO. And if you've been interested in what we've talked about today, uh, visit radar.cloudflare.com, where you can dig into the particular service you're interested in, the network you're interested in, the country you're in, whatever you want to know about how the internet is trending, and what we're doing with it.
BP Very cool. We’ll put that link in the show notes. Alright everybody, thanks for listening, and we will chat with you soon.
[outro music plays]