Dor Laor on ScyllaDB’s High-Performance Database, AI Trends, and the Future of Open Source | AWS re:Invent 2025

Hey, everyone. Welcome back here to our continuing coverage of AWS Reinvent. You know, we don't do every video interview live at reinvent because there's embargoes, there's other considerations.

And so this is one of the videos we recorded at, uh, reinvent in Las Vegas, and we're bringing to you now, just a few days later. I want to introduce you to my friend Dore Laur. Do is, uh, this CEO, I think founder of cid.

Yeah, Yeah. Co-founder, Co-founder of, of CID db. I got help.

We all need help. Do's been armed with me on Techstrong TV for years and years, but it, it's not often I get to see him. He's, of course, in Israel.

Uh, we were supposed to be in Israel right now, but we're not, uh, for Cyber Week, and it just didn't come together enough. But Dordt, it's great to see you here in person. It's great to have you.

Thanks. Thanks for hosting me. It's a pleasure.

So, let, let's start with this, though. Not everyone has seen you on text truck tv. We, you know, we're not, let's face it, we're not CNN or any of those, but yes.

Give people a little bit of your journey to, to founding, uh, Cilla. Sure. Um, so I'm a technical founder.

Uh, I have a roots in computer science, and, uh, initially in my career, I went to work for a terabit router company That early days tried to take over Cisco's core business in, in, in 2000, eh, the bubble burst. So we didn't work that much, but we did have a fabulous product and a drop in replacement for Cisco CLI, I'll, I'll come later on with more for the importance of, uh, drop in replacements in products. Mm-hmm.

Um, and later on I did something with Blade Centers, and then I joined the company, a startup company where I met my existing co-founder ti and my, uh, existing, uh, chairman who was, uh, the CEO back then. Uh, that setup had had to pivot three times. This is where, uh, I learned how to pivot Uhhuh.

The last pivot, we, uh, came up with the KVM hypervisor, so to, uh, renovate around the new hypervisor, a new approach that, that was the KVM. It worked really well, and Red Hat acquired the company. We, uh, spent their four years, uh, improving KVM and also the Linux Colonel, and I'm a big fan of it.

And, uh, afterwards we wanted always to have our own startup. So we, we left Red Hat and opened this company. Uh, originally, uh, it wasn't around databases because we had a lots of virtualization experience, so mm-hmm.

We started with, uh, an operating system that should have bit, uh, beaten Linux in, in, uh, virtualized workloads. Uh, the OS exists, uh, still today. And I met a customer yesterday who runs Sila and knows us because of that s uh, 'cause of that os Yeah.

If you don't mind, what os was this? It's called, uh, OS V, it's a unicorn. Oh, Okay.

Sure. Um, They had their moment in the sun. Yeah.

Uh, the Docker kind of sucked all of the air from the room when we around when we launched, but, uh, this is where we, we were familiar with other databases. We, we want to show, uh, the gains when other databases run on top of our OS to be faster than Linux. And we managed to accelerate Redis by 70% because we loaded the application into the kernel space was faster.

When we did the same with Cassandra, the performance didn't change much. We realized that the overhead of Cassandra, uh, is itself and, and not, and if you replace it with a fast os it, it doesn't change it. Uh, so we said, oh, that's can be a good idea for a pivot, because we didn't get enough traction.

And with why, once we rewrite Cassandra from scratch, keeping the compatibility like the Cisco days, uh, also like the KVM days, it's also about compatibility, uh, with other things. Um, and we rewrote Cassandra from scratch. That's what cila DB does.

Uh, it's also, uh, nowadays compatible with Dynamo db. It's a drop in replacement, and it's a standalone database that can run the biggest, most scalable workloads in the world. I love it.

What a great story, huh? Mm. And it's also, uh, you know, for, for geeks, right?

You're, you're, you're a geek person. I'm a geek person. A lot of the people out here are, we do this.

I mean, it's nice to be able to make a living doing it, but we'd also do it because we love Yeah, absolutely. Playing with this stuff, and, and this is a great story where your passion led you to, to doing this. Um, it's been now how long with s it's got six years, seven years, eight years, how long?

Mm-hmm. Uh, now it's, uh, it's more than 10 years, not 10. Yeah.

Even, uh, our 11th year. Really? Yeah.

That's, you know, what, and that's something also, quite frankly, to be proud of, right? Mm-hmm. Because what do they say?

The average company, if you make a past three years mm-hmm. It's a big accomplishment. So it, it's, it's all obviously here.

Um, now talk to me a little bit about how people engage with Cilla, right? There's open source parts of it, there's commercial parts of it for people out there saying, you know, we're always looking for better performance, better bang for the buck. What, how, how did they kind of jump into silla?

Um, so, uh, we started, we were big open source fans. Uh, we, we started with Open Source, actually, uh, a year ago. We changed the license, I remember to source available mm-hmm.

At the time, a year ago. I, I was just sitting here. Um, so it's source available.

We, we do have projects which are, uh, open source, like our source, what Even source available. Let me ask you a question. In the year you did that, how many people have asked for the source?

Um, so PE people do appreciate the, the, That it's available, The source, but It, it, this is, but this is something, look, I've been an open source too for 25 years. The fact of the matter is, 99% of the people never look at the source code or make a change to it. Not, maybe not.

99, 90 8% of the people never look at the source code, never make a change, you know? And, and so what they really want is free, Uh, yeah. People like free.

And, and we, we have, uh, a freemium offering right now. We're, uh, now it's source available. It's allows us to, uh, allow people to look at the source and, and also have the, uh, comfortability that the source is available for virus cases, uh, for future con continuity.

Uh, but, and we have some control to say, okay, up to this, uh, level, it's free and beyond that level, you need to pay because we are here 11 years on the road. And, and it's a business, right? Someone's gotta keep the lights on.

I, I agree with you, But I, I, I do understand people, uh, who are passionate about, uh, the source code. And, and there's a lots of, uh, small things and small changes where things matter. And, and we have, uh, open source, like, like our core engine, it's called csar.

Uh, it is open source, and its license is not a GPL, its license is, uh, uh, Apache because it's important for, for people to use it within their products. And that's why we haven't selected, there's a, a ton of no Ly Changes. Look, I, you know, one of the nice things that I've seen happen in the open source community over the, as I said, 20, 25 years I'm involved, is that most users recognize that though, open source may be free, someone's working on this.

Mm-hmm. Someone's entitled to get paid for their time and their effort and everything else. They may, they may quibble with how much mm-hmm.

But you, you know, it, it's ludicrous to think that people are gonna volunteer this outta the pure love and, and not make a living, you know, not be compensated for it. So I think that's been a positive development overall in the open source space. Mm-hmm.

Right. It used to be, oh, you know, you're looking, you're in it for the money. Everyone's in it for the money.

We have to, to keep the lights on. We've gotta feed our families. But, you know, it's just, it's a fact of life.

I mean, and if you don't wanna recognize that because you're some sort of, you know, like open source zealot mm-hmm. Free as in freedom and free as in beer, don't use the product. What can I tell you?

And, uh, having, uh, paying users allow us to invest back in the product. Absolutely. It makes the product better, Product better.

Um, so that's primarily what we do. And It's a flywheel As a, as a vendor that, uh, used to, uh, release both open source releases and also, uh, gated product releases. You double the amount of releases.

I was just gonna say, what a pain in the Yeah. You know what that is A hundred percent. I, I agree with you.

So there's, but there is a freemium version. You can go check it out, play with it. If you do wanna look at source code, and that's your thing, it's available to you as well.

Um, Dora, let's talk reinvent here. You guys are here. It's been an interesting kinda reinvent.

You know, we, when I, I just finished writing an article when I first got here Monday, and I looked at the keynote, you know, agendas and everything. They gave us a press preview. It was obvious it was all agent AI all the time, right?

It was all about ai. But over the course of two, three days that I spoke to people and saw things and walked around, see a lot of news about DevOps, cloud native platform engineering databases, hardware, hardware is AI stuff too. But hardware, um, you know, it, I maybe didn't hear as much as we normally hear about, like things like S3 or serverless or Lambda or these kinds of things.

But the geeks are still here, the developers are still here. The ops, the DevOps folks are still here in force. What have you seen?

Um, so AWS is, uh, a giant, yeah. E even more, more than that. Um, and nowadays they do innovation across, uh, across the year, not just them, also their competition.

They, they have to, um, so our announcement, I think that they're not holding the announcement just for, uh, this event, uh, recently they released, uh, new Graviton instances. Yeah. Graviton five is coming.

Yeah. And, and then in the, and the GRA Graviton four was released, right? And, uh, we are, we measured graviton four with cila db.

And, uh, it, it offered fantastic, uh, performance and that translate to better TCO. So for us, it's, it's super, that's exactly what we need. Um, so there, there's a lot of, uh, gradual improvements always on all of these products.

Yeah. Um, so it's for, for, uh, for, for, I'm, I'm pleased for that. It's, it's good enough for us.

What about, now I know you're exhibiting, what about like, you know, traffic at the booth, conversations with people? What are you hearing? Uh, well, there, there's, uh, no shortage of, uh, of traffic at the booth or traffic, uh, here in Vegas.

Uh, regarding, um, the entire AWS and, and the ecosystem, uh, it, it's mostly about, about ai. Like, uh, uh, we, we see that a surge in AI use cases. Uh, now about half of the use cases are a directly related to AI Ins, Cilla, Uh, ins.

Cilla. Yeah. In, in.

So explain that to me. What, what's the use case there? Um, we can pl split it to, uh, three categories.

Uh, one category is, uh, that we're part of the AI stack. And, and during the, uh, training and also the, uh, uh, serving processes, uh, the, the stack need to just access a tone of objects and, uh, need the fast database for it. It's part of the AI stack without doing anything, uh, special for it.

Uh, uh, distributed databases, easing demand for high workloads, and, and tho those are high, very high workloads. Sure. Uh, and, and c can be, uh, part of the big LLM uh, companies, or can be a smaller, much smaller company that started start their AR journey.

That's number one. Uh, number two is a feature store. Uh, feature store is more of, uh, machine learning, but it's, it's part of AI still.

And, uh, feature store allows people to classify, uh, users or, or sometimes agents, uh, automatically. So it can provide recommendations for, uh, e-commerce, for, uh, fraud cases in, in variety of other cases. And we're, we're big in, uh, feature store case and, and feature store needs.

Uh, a fast database too, to quickly come up with, uh, to, uh, classification that, uh, you as a user was selected and, and what's appropriate for you as a user either to watch on TV or to get an ad, et cetera. I love it. So this is the second one.

And the third one is, uh, a vector search, um, to, to do LLM on your private data set set. Uh, that's why, uh, the, this whole category of, uh, a rug Right. Was rag with vector database.

Exactly. So, uh, we added, uh, a vector search, uh, ourself. And we, we already have a, a beta that receives lots of interest.

And, uh, we, we are going to this month in December, uh, go live with the general availability of our, uh, rag, uh, vector search source. Really? Yeah.

That's fantastic. So in essence, they could uses, or is their vector database then mm-hmm. They're creating small language models or, or Yeah.

The rag stuff that's gotta be big. No, Yeah, that's, uh, fantastic. Our, uh, eh, vector search is the most scalable.

We can easily run a model with a billion, uh, objects. Uh, very few, uh, vendors can even get to a billion. And we can do that with hundreds of thousands of requests per second.

So we, we scale, uh, to, to very high numbers. And if, uh, people have a lower medium demand too, like, uh, most will have a model of, uh, 10 million or a hundred million objects, then we can give, uh, the best latency and, and also very low price point. That's fantastic.

Look, there's a lot of people saying that we've scraped all there is to scrape for these LLMs and that, you know, get, making generative AI or even agent AI better by increasing the LLM and the data we have to train is, is diminishing returns. And that the way to go is maybe SLMs more rag, you know, uh, well, there's some people who say, we need to go away from LLMs all together and go to this world model and stuff like that. Mm-hmm.

Um, but certainly, I, I believe there's gonna be a lot of activity in, in the SLM rag kind of space. And, and not only that, because as we develop AI for specific use cases, I don't need the whole world of the internet. I just need, especially if it's my own proprietary information, right.

And I don't wanna put that out up there. I want it right here. Just, and so I, I think that's a huge business for you guys.

Yeah. Congratulations. Thanks.

Uh, it's, it's, uh, the market demand. Yeah. Yeah.

Well, no, that Is, it's not just an opportunity. It's also a defensive move. Because if we won't do it, then uh, customers will go elsewhere.

Uh, to, to be frank. And yeah, the, the fact that, uh, people would expect, uh, all of the ease of use of LLM on the public dataset on the internet, they expect to have the same when they come to every vendor. And to ask if free text search, uh, your questions in, in one liner, and get immediately the best results without diving into a very complicated ui, that's a power of LLM.

And sometimes it won't be people, but it'll be agents, right. Uh, that come and, and automate and, and get the queries automated. So that begs the question, is there a an MCP server in your future, Uh, in the future?

Absolutely. Yes. All right.

Hey, let's fast forward past AWS for a second. People are watching this after the, after the show. Anyway.

You guys have some new announcements that you're previewing here. Mm-hmm. Share, if you don't mind a Little bit.

Thank you, uh, for the opportunity. So, um, uh, we'll also move, uh, from beta to general availability. Our X Cloud, uh, uh, a managed platform.

Uh, X Cloud is, is, uh, the new generation of our core database with, uh, database as a service management consumption. Uh, the unique thing about it is, uh, our new core architecture, which is called tablets. It's way, way more elastic than any other database or even infrastructure in the industry.

Uh, we, we were okay with regard to, uh, the speed of, uh, increasing the cluster, scaling out, and then scaling in. We were, before this technology were, we were okay, like, like, uh, an average vendor, but there was demand to do it much faster. And frankly, we also compete with DynamoDB.

We're drop in replacement, and DynamoDB, uh, was the first no SQL database. And, uh, up to this change was the, the best in the industry. You can easily scale up and down, uh, very easily.

And, and if your workloads changes throughout the day, uh, then you can, uh, instead of paying for the peak consumption all the time, you can just have the workload follow, uh, uh, the work, the workload should follow the usage, right? Dynamically. So that's exactly what, uh, X cloud is.

Uh, we have a, the technology based on components called tablets. We break the gigantic database of, uh, a petabyte of data to five gigabytes chunks. Right.

And we can move them around super quickly. Uh, we, we can also even, uh, it allows us, uh, both to scale super fast. We, we can increase capacity, quadruple it in 10 minutes.

Mm-hmm. So you can go from, uh, 500 K to 2 million operation per second in 10 minutes, But could you go back to 500 K and 10 more? And that's right.

So, Because sometimes with these things, it's like blowing up a balloon. Mm-hmm. You know what I mean?

It never goes back to the size it was before you blew it up. So we, we can, it, it's not, it, it's, it's, uh, indeed complicated. Yeah.

But, but we can also go back and, and shrink and, and that's the user workload that, uh, goes, comes and goes, whether it's a Black Friday or, or on a daily manner. Uh, so, so that, that's a big improvements. Uh, and, and big TCO improvements and, and usability improvement.

Sure. Uh, also, it's, it's, it's pretty unique. Uh, we have a short per quart, uh, engine.

So let's say if you have, uh, a machine with, uh, 32 cores, we, we'll have 32 independent threads in the server. Wow. Uh, if you have a 64 machine, then we will have 64 threads, uh, in, in engines within that machine, and it'll perform twice as good.

32. Now, let's say if you have a 64 way machine, uh, but actually you need, uh, um, 66, uh, threads and you have 64, and now would you buy another machine for 64? It's, it's expensive, right?

So instead we can mix and match, and we can have 1 64 machine together with, uh, a tiny two VCPU machine next to each other because of the flexibility and the hard. So it's Real Distribution and the sharding, we, we can combine the two. Haven't seen any other vendor can do that.

No. And what the user receive is efficiency. Uh, they have exactly what they need.

They don't need to buy excessive large SER servers, which are expensive on AWS, uh, They're expensive everywhere. It's not just AWS but really what we're talking about, here's almost like a finops play, right? Because that's, I think that's where we are, especially in cloud usage, right?

Look, we're talking about spending $5 trillion on data center AI factories, but the fact of the matter is, when I talk to people, they say, I wanna get control of my cloud bill. Hmm. I wanna reduce, I wanna be more efficient in my use of these resources.

And, and that's why I made the joke with the balloon blowing. That's pretty much how the cloud is, right? It never seems to go back down.

People, they want that ability to have insight to turn that dial, and they want the ability to say, how can they do this more efficiently? Mm-hmm. Yep.

And our customer success team works with customers. And if we both see, let's say you sometimes utilization people can check their database, how much it, it's loaded on an average basis. Most databases are, are not that loaded.

Uh, on a, when I'm not talking about the spike, I'm talking about normal, uh, day usage overnight, it can be 10%, uh, or 20% utilized and you pay for the entire thing. But that was always the pro, that was the promise of the clout. That elasticity was a up and down thing.

Yeah. It wound up being more of an up thing all the time. But it's good to know that's there.

So this available, well, by the time people are reading this, it'll, or excuse me, by the time people see this, it'll be available. It, it's, uh, today, uh, a avail dated to, uh, a WS conference available as beta and, uh, the time people see it available as general availability. Excellent.

Good stuff. What else from Sila? Um, so it's mostly this.

We, we do have, uh, lots of, uh, things that we develop like tiered storage mm-hmm. Uh, in, in other technology to, uh, reduce the bill. Uh, normally we use NVME for fast storage, fast performance, and it's also relatively cheap co compared to different alternatives of, uh, of storage.

But, uh, SS three is cheaper. The problem with S3 is that latency is prohibitive big. It's a 50 millisecond, 100 milliseconds.

Uh, and with the storage, uh, we can keep the hot data on facet and VM e and automatically move the cool data to S3 and come, come up with, uh, a good solution. 'cause sometimes you keep, let's say 30 days of, uh, of history on, on, on Sila in the NVME, but you'd like to keep one year of data and, and access it through the same API and not develop a new access for it. So this allows users to, uh, have one API and, uh, a very cost effective solution.

I love it. Good stuff. You know what, we didn't, we didn't even mention the website, URL for people.

Want to go find all this out on their own. Dig in a little deeper. What's the, what's the best URL to go to Dore?

Thanks. com. com.

Just as it says underneath is in his lower third. Alrighty, do. It was a pleasure seeing you.

Safe travels back home. We are wrapping up now. Again, you, you're seeing this after we were here at, uh, AWS reinvent, but it's part of our AWS reinvent coverage.

And if you need to find this back on, it'll be listed under the event coverage. But for now, this Alan Shimel for Textron tv. Thanks for joining.