Reddit CTO Chris Slowe: The Engineering Decisions That Built the Front Page of the Internet

Reddit primarily exists as a platform for criticizing Reddit, it took the scenic route to becoming a major part of the Internet. Tech debt is the sign of a not dying startup. There is still a lot of code in our stack which still has my git blame on it from a much earlier vintage of me and that guy didn't know what he was doing. So I am both proud and terrified that it's trust. It's passed the test of time.

Today, the CTO of Reddit and technical architect behind the sixth most visited website on the Internet shares the engineering decisions that have led the platform from a scrappy startup to a global giant, navigating massive rewrites and building for a deeply opinionated community of millions. We dig into the reality of scaling technical teams, why hope is not a strategy, and what it really takes to keep the front page of the Internet running for two decades. Chris, thank you for coming. I was telling you on the way up as we were coming that I trace back the start of my red journey to 2011 peak neuroplasticity. So I so I trace a lot.

Of my so you're saying we've corrupted an entire generation?

Because I remember I watched a talk and you said it's hard for me to take myself seriously and it might be because I grew up on Reddit. So I have a little bit of that as well. And then founding engineering current CTO of Reddit Chris joined Reddit in its early days back in 2005 as the first hire, earning a PhD in physics at Harvard, architect behind some of Reddit's most foundational systems, including the hotness ranking algorithm and the early anti spam infrastructure that helped keep the platform usable as it scaled and then left to work with Steve on another startup, the CEO, the current CEO of Reddit and how did Verplexy Deep Research do?

Did pretty good. It was actually 2016 but close enough and at this point that's who gives 8 or 9 years either way. But no overall good LM summary there. I think there might be something to this technology. I'm not sure getting a hint it's not just a fad.

Well so one of the things that was the most interesting for me to put myself into your shoes as I was thinking about some of the things I want to talk to you about in this is thinking about the redesign of Reddit and at The Time Top 10 Most Used Website in the world. And so I think something that would be an interesting place to start is in retrospect, any words of wisdom, things For a team who's about to go through a massive rollout.

Well, okay, to start with, I'll say what I always say in situations where a redesign is involved. Congratulate. It'll be fun and harrowing and terrifying. I also keep in mind that Reddit took a very. It took the scenic route to becoming a major part of the Internet. I remember going back, like, way back in the day. I'm sure no one here remembers originally there was these rankings called Alexa predating the Amazon service, where it was just effectively the rankings of all of the sites on the Internet. And in the early days, we used to track our Alexa ranking and just kind of watch us break into new tiers. And I just remember roughly the same summer that Reddit launched, YouTube also launched. And I just remember seared of my memory is like our lovely, like we're making progress. And then you overlay, overlay YouTube and it's just like vertical line and just appears right in the middle of it. So there's always somebody who's going to grow faster. But compounding growth over 20 years actually gets you somewhere is the lesson we learned for the specific the redesign. Oh, man. I have lots of opinions. Okay, so I'd say start with where I've seen things go terribly, horribly wrong. The discussion usually starts with, well, while we're in there. And then usually you make a decision that starts to push you towards overreaching how much you're going to do in the redesign. So what's good to try to do is just try to separate out, I don't know the scope of redesign, but try to separate out as much as you can. UI changes and user experience changes from drastic backend RE architectures. Now they're going to be coupled because you can't just usually slap a code of paint on the front end. And of course, modern practices as they are, everyone likes to build web apps in such a way that they are absolutely transparent from server to client. And that's really great from a velocity standpoint, but it's really tough from like a separation to concern standpoint. Make sure you can revert, honestly, would be my starting point. Or make sure that you're at least starting with small scale, ramping up. The way that we did the redesign back in 18, and actually the way we've been doing it ever since, when we do like a major rebuild is it's a little bit beyond just doing simple A B testing. It's actually starting to identify user cohorts that are going to get the new experience, which is difficult in Reddit because, you know, Reddit primarily exists as a platform for criticizing Reddit. So you can't just secretly launch things and have people be like, oh, this is a great novel experience. I love this so much. Instead, it's like, what is Reddit doing? But what we did was for the redesign, we actually broke it into, like, we did a private beta where we had users involved in playing with the beta, and then as we had success with the beta, we pulled in more users. And then the nice thing about that is you get a nice little flywheel where it's not like you're not trying to educate the users, the users could educate one another. And then also you get some advocacy for the actual thing. So everyone gets a little bit of like, the people who start to use it start to realize that it's not horrible and you're not trying to destroy the product they love and they can actually advocate for it as much as you can. Wanna go deeper?

Talk to Chris's digital mind@delphi AI chrisslow Another thing that I've also heard you speak a lot about is processes and systems start to become culture. They're a lover for culture. And I'm wondering what were some of the things that you found that started to break once you got to that level and the corrective measures you took.

So I would say that one of my learnings over the years has been you do need to. When you set up a startup, you have to start by minting at least a starter set of values that you can share with the team. And you can think of values as being a couple of different things. So one way I like to think of them is they're kind of like verbal shortcuts for behaviors. So they're a good way to, you know, when you have a discussion or a debate or an argument, if you can resolve that by invoking a value, it saves everyone a whole bunch of time. And it kind of immediately gives you alignment, or at least grudging alignment of like, okay, that's how we operate. And therefore we're going to kind of move on from this point. It also means that as the company gets larger and you can no longer rely on just the simple social graph of the people in the company that you can still have some notion of almost common culture to be able to align to, there's a whole set of these dumb, I don't know if you know dumb bar numbers, but there's a whole set of them. It's actually not just one. There's actually Tiers effectively. And the tiers break down the size of the group relative to the overall intimacy level you can have with them. Right. And so when you're talking about a room this big, which is around a dozen people or so, Right. This is effectively a family, like to be kind of, you know, and not in a condescending business. We're all a business family. But like, you know, effectively you can, you can, you can know each other's stories, each other's names. You have a whole bunch of backstory, a bunch of context. And so if somebody, somebody rolls in and looks really pissed off, you probably can make some guesses or you want to check in with them because you're actually on like a much more familiar relationship. When you get up to about 50 people, well, that's the point where you start being able to. You lose track of individual stories and individual people. You can keep track of names, but from like 50 to 150, you're kind of at the range of casual acquaintance. You might know them, you probably run into each other in the kitchen. You kind of get to say hi, but you don't have any shared context. You might even know what. You probably don't even know what each other's doing. And then when you get past 150, it's like all bets are off. It's like you're effectively strangers. You run into somebody, one of the startups problems, you end up at the 150, you start having these problems of why is that team messing with our stuff? And the answer is that teen doesn't even know you exist, let alone messing with their stuff. They just found a service that they had to change and be like, you run that service and yeah, they should have reached out to you, but did they reach out to you? Did you answer them when they reached out to you? And so back to the point of culture is now to start to set up some, just some habits and some behaviors that are transferable across the entire group, even when you don't have any shared context beyond that.

And so one of the ones you guys have, remember the human?

Yes. The oldest one, the one that actually stuck around for a very long time.

I was going to ask what was.

The origin of that. So this is mostly. So Steve does a sort of like usually once every two or three years, he goes through and does a refresh on our values. And remember the human was. So when he came back in 2015, first thing he did. And one of the things we learned in our intermediate startup was like this whole Thing about, oh, values are important. So he came back and he basically built the first set of values and that was one of them. And the intent was to be, I think, cover a bunch of things. So on the one hand it's remember that you were building a product and there are humans on the other side of that screen and therefore your work is affecting their life. And then third thing is internally it's nice because it's like, remember, assume good intent. There's other people here. You're not just acting alone, you have to work together as a team. And so it's actually been the one value that's had the most broadly applicable use, which has actually helped a lot. We've had other values that I think that over the years have varied. And so actually when we were a startup that was a much larger than this size, we had one that was everyone does the dishes. And the intent of that was both literally, hey, we all share a kitchen, please don't leave dishes in the sink. Because we're effectively roommates here when we're working during the day. But it also covered on the way through engineering, it's like, yeah, don't leave shit around. If you're going to go clean something up, finish cleaning it up. Don't leave a mess for somebody else to have to come along and clean up. And of course that value actually that lasted for a while and then we ended up sunsetting it. And in part it's because the parts you want become part of the culture and they become automatic. And in part it's because, well, at some point you actually just need to hire somebody whose job is to do the dishes. Like you don't, you don't assume, you don't, you know, tragedy of the commons is a real thing. It works. Unless you have like some sort of social pressure to make sure nobody's like, you know, some people are not leaving their stuff in the sink. You probably need to hire a janitor sometimes.

And so one of the things also.

That.

Sticks around for a long time, we're talking about cleaning up messes and not being a strategy, is tech debt that you incurred. And I'm wondering, I'm sure you have some interesting stories. What has been the longest running piece of tech debt? Maybe on this even now, or you already have things that are coming to mind.

Okay, so before I answer that question directly, I will say that what I've learned is that tech debt is like any other sort of debt, right? You take out debt because you want to have a short term speed up and you just know that with those choices you've made, you have to pay it down at some point in the future and it compounds, so it's going to stack up. You all know this. I think I've also learned that when tech debt, like tech debt is the sign of a not dying startup, if tech debt is your biggest problem, you probably have gotten far enough past the product market fit to have other problems to deal with. Or at least your existential problems are handled right. You probably have a sales cycle, you probably have a product that's working, you're now working on like maybe iterations and new versions. It. It means that you're not just going to do a pivot where you're just going to rip the band aid off and greenfield a new product, right? Oh man. I mean we have. Okay, so it's actually getting really good right now. But we are about in year six of the Decommission, the Monolith project, which I think everyone at some point goes through, at least from a certain genre of startups when you. Of course we do. So you only have 18 people, right? So you only need one code base, right? You don't need microservices when you don't have teams to break the micro across. Right? You don't need separation of concerns, you just build all in the same pile. And the nice thing about monoliths are they're easy to deploy because they're effectively global. They're also easy to recreate in red green if you want to do experiments because you can just have two partitions. Traffic routing straightforward. Like hundreds of reasons why monoliths are great. Monoliths are annoying when you have more than about 50 people and they get impossible when you get past 150 just so you can't scale up. So everyone goes through this of having to break off services. The problem is that the monolith is just a pile of business logic over time. So my point was there is still a lot of code in our stack which still has my git blame on it from a much earlier vintage of me and that guy didn't know what he was doing. So I am both proud and terrified that it's passed the test of time. And in fact my running joke has been like, I cannot leave Reddit until there is no more code that I've minted in the stack. My legacy should be net zero diff. But no, that's probably some like disentangling our very old. I should mention Python 2 Monolith has been a big deal and then on top of it all, we had a couple of cases of intermediate attempts at building. We actually had a service called the Gateway, which was a way to sort of put an API break on top of our monolith and start to reroute traffic properly. And that thing turned into another monolith that was kind of like co dependent with it. So we've definitely fallen down the. You know, we've made. Mistakes were made. That's what I'm saying. But it is a game of just kind of like for that kind of stuff, you have to actually have teams dedicated to chipping away and building the right thing. It can't be a side project.

Yeah, well, so one of the super interesting push and pulls that Reddit has with this community is that people get super fixated on the way that it currently is. And so it is currently perfect.

I don't know. You knew that. Any change, if it's perfect, any change from perfection, it's not perfect.

You're trying to destroy it. It's very straightforward when you look at it that way.

Also, why did you change that feature when there's another feature much more near and D to me than I want?

You should just think about that.

Yeah, clearly you have to think about it. It's so clear when you say it like that.

And how, how are you guys straddling the. That dichotomy where there's so much opportunity right now to make features that are especially driven by AI. Like I texted you about the AI search feature that you guys released, Credit answers, and there's a lot of opportunity for that, for new features, but also not trying to alienate the strong active user base you guys have.

Well, I think the nice thing what specifically answers is that it is a, it's a complement to existing flows. Right. It's not in and of itself a destination. It's a way to think, to rethink how could search be on a platform like Reddit where there's tons of good content and the discovery of that content is a difficult problem. And so we're taking the bend of effectively not just providing you with an LLM summary, but actually citations included as first class, easy way to get to the content that actually shows the conversations. I think that the balancing act, and this has been something that we've not nailed ever, I think very few companies have nailed it is finding the balance between doing experimental features and making sure that you're not just creating a ton of debt in the process of launching all those little features. If you think about if you have a thousand random experiments happening on your platform. Well, first off, in that situation you don't even have a common experience for the users. There's a whole bunch of variants that you're seeing, so it's very hard to talk about what's the average experience. And then most experiments are null results. If you're lucky, that's the net result. The accumulation of a ton of little experiments can actually be like all those little epsilons can add up and become something actually non trivial and negative because you're getting kind of a weird experience for every user. So I think what we've been trying to focus on in the most recent years has been, well actually let's just get back to good old fashioned, like what's the craftsmanship version of this look like? What's the first principles? What are the experiences we want to have? Where are we actually building things that help users and help them find what they want or engage and enjoy? And when are we just kind of like faffing about because it feels like we should be fiddling with pixels and making sure we kind of don't let that part win. But it's not easy. I mean, it's like, it's because like, you know, every. I mean the other hard part is that with a sufficiently advanced experimental toolkit, you can convince yourself you've created a positive roi, right? Because you can always look in the experiment, in the experiment and say like, oh man, this moved like D2 comment generation by 0.1%. It's a win. It's like, were you trying to move that? It's like, did you just happen to look through and look for something green? Like what was exactly your plan? Well, I mean one thing, one choice we made is we still have like old Reddit still exists, it's still up and it's consistent, sustained engagement from effectively the same group has been using it for 10 years running now. Because we actually, you know, we did our last major redesign was in 2018 and then we just did another web redesign in the last couple of years and deprecated that one. So it's actually outlasted two or three other web platforms in its process. There is something to be said about like, you know, if it's not hurting anybody, keep a separate version. But of course, you know, there is going to be overhead in maintaining that separate version.

So another thing that AI, outside of the opportunity to produce a bunch of new products is changing is probably people are putting a lot of AI generated content onto Reddit. And so one of the things that I'm interested in is how models. Have you played with VO3?

Yes. That's pretty cool. I think the hard part right now is every six months there's a new model that comes out that's like, wow, this is the best thing ever. And it just kind of keeps coming. So it's really impressive Also, I'd say, yeah, the posts using VO3 on Reddit have been really hilarious for quite a while now.

I mean, someone sent me a tweet of Moses live streaming the exodus with VO3. We're going to talk to the Pharaoh, see what he says. And so now I'm super excited. It'll be a subreddit, probably of historical live streams using models. And so there's the very positive creative aspect that I'm super excited for because there's insane amount of creativity that comes out of it. But what do you think is the best way to. Or a way to help people know when they're looking at AI generated things if it isn't intended like, it's not in the AI movie subreddit.

So we've thought a lot about this one and it's actually been like, it's definitely been also years in the making. Right? Like deepfakes as a concept go back a good five years now, give or take. The models are just getting better with time. I do feel like there's a couple of elements here. So I think one thing is there is like, there's just the good old fashioned cultural element of like, humans are good at calling bullshit on things if you let them. And you go back 20 years and the same arguments were being had with Photoshop. It's like, well, how do we know any photo is real if you can Photoshop it? And now it's just kind of like, well, everyone assumes photos are face tuned and slightly tweaked and just kind of like you kind of look at it and be like, yeah, should look like that. Those eyebrows don't look right. But I think with things like with the whole bevy of tools from LLMs, it's just an extremely powerful tool to create extreme creativity. And at the same time, you can use the tools to help you detect the tools in any other arms race that you would see anywhere. I think for us, we try to think more about the kind of incentives and the motivations for why you would do that in the first place and work back from that as opposed to just trying to directly say, let's go through and try to target label everything that we think might be from an LLM like the one that came up actually recently that I was actually kind of amused at was there's an article about how, I guess ChatGPT tends to use EM dashes a lot, right? And so EM dashes are everywhere now all of a sudden. Well, first off, I also happen to use EM dashes because if you just type in dash dash, it turns into an EM dash. And basically every editor known to man. So the fact that ChatGPT uses it more preferentially was kind of a funny thing. Second, it's like, well, you could also figure out like, okay, a lot of people Also now use LLMs to help them write. So if I'm going to take something and basically copy it out of an LLM, it doesn't mean that I wasn't writing it. It just means I was using a different tool than I have in the past. I didn't go to my editor and use all my thumbs to type the answer. I actually wanted to write something cohesive and therefore I'm doing my normal editing loop. And then I think for the actual abuse cases here, we think of it in terms of the same way we think about honestly, historical spam problems, right? Like, why are you trying to push content on a platform? Well, you're probably trying to either run an influence campaign or you're trying to make a bunch of money. And those two things are not necessarily separate. They're almost the same thing. In a lot of times, those kinds of behaviors are a tale as old as time on a social platform. And so it does fall into the category of looking for what we would just label as inauthentic behavior, or the class of spam problems where it's like you're trying to build a reputation so you can deploy your payload later and make a bunch of money off of it and then move on, right? That can come from signals beyond the actual literal text and the literal image that's being posted. And then I think the last leg of this tool here is we work via communities. So to some extent, I think watching communities wrestle with this particular, you can call it a problem, you can call it an opportunity, and figure out where they align. Like, you see a whole bunch of our communities who'll basically be like, none of that AI slap, only hand drawn artisanal drawings, right? And then the flip side of that is you have entire communities dedicated to specific tools showing how crazy you can build Moses crossing the Red Sea live streaming. And so we want to, again, we want to attack the behaviors and the outcomes and what the communities wrestle with the how does that tool fit into the ecosystem of what other tools they have available?

Very interesting. I haven't thought about categorizing it in. I mean, two things. One in the same vein as just traditional spam and then also thinking about going back to the incentives instead of trying to attack the behavior. I was going to say maybe this is the on ramp for crypto. And we put what are we buying and selling now? And we put watermarks that are cryptographically verified in every photo. And then the platform Reddit make sure that the picture doesn't have.

And it's all backed by a zero knowledge proof somewhere on a blockchain.

Okay, exactly.

So when do you want to start?

We should.

You want to pivot? I'm joking. No, I think so. I do think that for a platform like Reddit, there's a bunch of text, it's just a beer in a box. Right. And so the point is the conversation. And so the point is to create a system where there is trust in the. There's at least trust in the underlying community, there's trust in the interaction, and then we're able as a platform to kind of keep the jerks at bay for all. Definition of jerk. Right. Or we provide the community with tools to keep the jerks at bay even better. Right. One of the things that is actually novel about Reddit is we do have an ecosystem of self identifying bots who run around and try to be helpful. And that goes back well before the existence of LLMs. This is like 15, 20 years old. As soon as we opened up our API, all of a sudden people started writing agents against it. And all the way from the metric to English and back and forth bot that existence runs around. Or there's one that'll identify when you've accidentally written a haiku in your comment. I think that's always kind of funny. But the fact that those are self identified means that there's trust in the system. Because you know you're dealing with a bot where the line gets kind of blurry is like, are you like, dude, are you real? Is when you have to start getting into the thinking about the systems and the incentives and how do you make sure that there is policing for that kind of behavior.

I mean, something that I am continually blown away by with Reddit is the amount of time that people spend making very high quality creative posts that are well researched and the amount of the volume of quality content that people put on it. There's been a couple times in My life where I was about to, I was preparing for a long trip or something else and people would have just insane guides. And then the other thing is in the comments, the self policing where people will say, you know, oh, that was a mean comment, or this is AI generated and just you'll open the comments and people are very, very passionate. And so that seem is, I think it's a very unique thing about Redis.

The funny thing or the funny discovery was we made a decision early on that we wanted to stick with the kind of not really anonymous, but this kind of pseudo anonymous setup of you create a username and you just kind of post a bunch of content and you're off to the races arguing over effectively stupid Internet points. Right. And the nice thing about that is it feels like the first principles, most of the social media, social platforms pivoted the other way and saying like, no, no, you should attach your name, you should have skin in the game, you'll behave better if it's associated with you. And I think we found it's actually the opposite is like people no longer care about their name being attached to things, but at the same time they do put a lot of effort into posing and posturing and influencing and trying to create Personas that are kind of false. Whereas on Reddit, since it is all synonymous, you don't really have any. You're not trying to prove anything. It's like you're going to be helpful or not. And yes, there is an abuse vector, but there's an abuse factor on every social platform because you don't have people. But the positives definitely outweigh the negatives. The positives are, yeah, you might just be some guy who happens to be a big fan of horticulture and some obscure thing. And there's a question that appears, it's like, oh, this is my jam. And you just write a long answer and you're not expecting anything out of that than just this kind of altruistic, I want to be altruistic or I want to feel good by being a part of this community, or I care about this community, I'm helping it. And that's a lovely virtuous cycle. It's also extremely difficult as a product to explain, by the way.

I have rapid questions and then we'll open it up to everyone else. So first thing is react versus lit.

Why? Oh, we have done both. Lit gets heavier the more people you have developing on it and it gets to be really hard to debug. There's something really nice about Server side rendering HTML. Okay. It's almost like the entire system was built on top of that as an idea.

Favorite ama, Donald Reddit.

Oh, man. I think that there was a tradition we had of, honestly, Bill Gates would come by like once a year and he always was just so funny and engaging. I think I can't not name the fact that at one point President Obama came and did an AMA that was a terrifying experience because that was definitely like, you know, harrowing. There's been some really good ones also. Like Keanu Reeves. Exactly what you'd expect. Just like a. Such a. Such a cool, chill dude. He's. Isn't he great? Like, he's like a. He's just. It's always fun to find. It's always fun to be reassured that there are people who are just awesome people. Like, there's no secret they're not just like, they're not secretly like breathing fire eating babies. They're just like, nope, he lives the life that he acts like he lives.

Craziest idea to come out of Crazy idea Friday during your redesign.

Oh, man, you're digging deep into the well on this one. I gotta go back and remember those periods. We had a lot of. We had a lot of fun ideas for projects that never really took off that we've had to since unwind. I think the one that actually still. I still believe could have worked was we tried to set up an entire. I think we've iterated on it since then. We had this idea of like, helping to set up like an event flow for communities that are around like fan bases of like, TV shows. So it's like, you know, automatically trigger the post when the, when the episode drops and kind of get that whole thing going. It's. It was that one. Was. That one was tricky to kind of land.

This, like, to the point of people who are super invested. This happens. The mega threads.

Yeah. Mega thread perhaps like the. We have to. We still have a problem right now that we have to do. Like, for a certain size of thread, they're hard to manage and actually the easiest thing to do is just have them auto roll new ones and like substitute them. But mega thread management is definitely like.

A. I would go out and only.

Say it's an underinvested area right now. It ends up being. The hard part is it's such a small slice of the overall traffic profile.

Yeah.

But when they're up, they're just like. So they're such a big deal.

Yeah. Because I found myself the. I found myself looking at Red every day during the Lakers playoff run because I'm a Lakers fan and they're always mega threads, you know, whenever they're playing or people are talking about the game. So the next one is worst architecture decision.

I'd rather. Oh, man. Okay, so going back in the wall here, I'm going to dis an entire probably technology. We decided really early on to move to Cassandra and this was like Cassandra version 0.6. So bad choices were made by going to sub integer version and we started using that to. Basically the quick version of it is we use it as a way to build not quite a data store, but almost like a persisted cache. Like most of Reddit ends up being like materialized views of content and you want to basically have that be fast and perky and updated and you can rebuild it from the canonical source, which is a database. And when originally we put it together, we had it running behind a cache and then we dropped a cache and we had to go straight. We had to go into Cassandra to do those queries and we discovered that it wasn't actually performant enough to run. So our longest downtime was like 24 hours back in 2010. And that was pretty brutal. Okay, Cassandra, we dug a big old hole in that one for a while too. And like we've, we've since, you know, we've, we've modernized since then. It's been like back and forth. We had to hold it now, like, I think that that was the. Actually the biggest mist. I can summarize the class of biggest mistakes we've made is, wow, look at this cool new shiny technology. It's going to solve all of our problems. It clearly has some really great maintainers and they know what they're doing and it's going to be great.

To that point, picked a beta version of an ORM wrapper adapted from Typescript to build our entire backend on that got sunsetted two months ago.

Oh yeah, okay. Yeah, yep, that sounds Boris. That's those. Them's the breaks. Them's the breaks. Oh, actually I thought we'd been better by the better dumb decision we made. Now that we mentioned Postgres really, really far back, we're going back into like first couple of years of Reddit. We had this random. We had this idea of building an entire version of Reddit that was, I shit you not, based on postgres triggers. So every time a vote would come in, it would recompute all stuff in the backend. It was almost like, you can imagine this machine was flicks Switches, flicking and like most things that are designed like that, it worked really well until it didn't. As soon as you scale past the ability to have enough compute to do that, it broke really spectacularly. That's really funny.

How do you guys keep your materialized views up to date?

Oh, man, it's layers and layers of CQRS models. So what we do is we have most of it is like Reddit is mostly not real time and it's mostly eventually consistent. And so there's just a lot of, like, you take an event, you put it on a queue somewhere and you fake it until it actually gets computed. So for that user, that vote happened immediately. And then actually what's happening is, you know, there's a whole bunch of. There's a whole Goldberg of things of like 20 years of technology that are processing that vote and making sure all views are updated properly. Best architecture decision using Postgres. No, I mean, seriously, it's been like rock solid for. Actually, I think I've just started hearing from my infrastructure team that they are looking around a little bit, but it's only because we have just scaled the shit out of Postgres over the years.

Is this still just Postgres all the way down?

Well, so it's like if you've gotten down to Postgres, something drastic has happened, right? There's not very many. The issue with Reddit is that we have an extremely long fat tail. And so at any given time there is probably requests that's not in the hot path. But of course the name of the game is to get as much in the hot path as possible, right? Like, you know, we have like terabytes and terabytes of cache everywhere. And then underneath the cache we have like an equivalent amount of almost like, you know, Cassandra, like persistent storage that you're not. You're not building things from scratch, you're actually fetching them like materialized. What is the cache largely these days? Redis, which is super solid. Cool, that's great. I think actually though, I'd say with the current now sunset kerfuffle about Open Source Redis, we have been looking at evalki as an option as well and trying to decide on which way that community is going to go.

Favorite subreddit.

This changes a lot of time. I think that I'll give examples from my own history. So they change based on my life, where I am in my life, among other things. And so I definitely am still a big subscriber to a bunch of makery kind of communities. I Got my start doing experimental physics. I love building things. I don't have access to machine shop. I have a 3D printer and a laser cutter and a bunch of tools. So I love all of the fix my print ender three functional print. All those are fun communities. I've got two kids I've got now 14 and a now 11 year old. When my youngest was going through his terrible tool twos, there was a community that happened to hit at the same time and it was called you. I'm a toddler. It was just good examples of just like we used to joke that he was like he would sort things into things that he could push over and things that would push him over and that was all. Never the twin shall meet. So yeah.

You were part of one of the first YC batches where Algram.

Still used to cook food. Like are there any like furries of fear of? Oh yeah. I mean like those are really fun. So this is like Reddit. Okay. So actually I was in that batch before I was in Reddit. I was actually in the first YC batch in another startup. So I already had a failed startup behind my belt when I joined Reddit. That first batch was 10 people or 10 startups working out of Cambridge out of an office there. And the weekly dinners were the way to kind of connect with the fellow co founders and basically can trade war stories. And it was a lot of fun just because at the time 2005 there just wasn't much of a startup ecosystem. The whole idea of Web2 was relatively new and so it was almost like everyone in that room was kind of learning it as you go. At that point it was hotly contested. Is Ruby a language we should try to learn? Do we go back to some php? So really just that camaraderie of just talking about building a company from scratch. Also this is a point where in that first batch there were very few companies that had products to show when that batch started. It was an idea pitch to start with more than anything else. So it really was that first summer with building the first startup. Oh yeah. And also Paul used to refer to the food as glop specifically because there's always some sort of like a rice with like a curry kind of thing. And therefore you get your tray of gruel and then go sit down. It's actually really. He's actually a really good cook.

I mean that batch was. You see a picture of it, the amount of talent density.

Well, with time, it's amazing. Yeah. So in that batch there was. Okay Let me try to remember the names of the companies then. So there was a company called Loopt started by a gentleman by the name of Sam Altman. There was a calendaring company by the name of Kiko that was started by Justin Kahn and Emma Cheer, later of Twitch. And then Aaron Schwartz was in that group doing Infogami, which later merged with Reddit for a time. So, yeah, a bunch of really smart people. All right, cool.

I think that's everything.

Thank you so much. Thank you.

Thanks for tuning in for personalized advice from Chris Head to Delphi AI Chris Slow and ask his digital mind.