What Is LTAP? Lakebase + Genie, Explained by Databricks CTO Matei Zaharia

0:04All right. So today we have a special guest Matei. Thank you so much to be here. So Matei, we don't need to introduce you. You created Apache Spark, CTO of Databricks, co-founder of Databricks, still professor at Berkeley. Uh you build on Unity catalog and Omnigent. I thanks for being with us today. Uh and thank you yourself also to be there.

0:26>> Yeah, thanks guys. Thanks for having me on. So DAIS was super exciting like 30 32,000 attendees in person almost 100 if you count people who los joined remotely 174 countries a lot of great announcement LTAP new architecture LTAP raiden with lakehouse RT of course AI which is uh your domain if we

0:49if we get if we don't include omnient the most exciting thing happening at Databricks right now >> yeah it's definitely been a great year for uh for launches and a lot of things coming together. I think there's two really big ones. I think the first one is Genie. It's it's gotten so uh good now at actually understanding like uh many types of business data and many

1:12types of questions. Uh in particular, you can add unstructured data too now like your documents in SharePoint or Google Drive or uh anything like that. And when you start to use that together with your business data that's in tables and you also, you know, it has the ability to create like charts and and and reports and stuff for you. Um, it becomes very very powerful co-workers.

1:35So, I'm really excited about that. I use that all the time like just in my normal job to figure out what's going on um and uh to make plans and and things. Um and uh it's great to get it into more people's hands and uh and you know make it really work uh well at the different domains. Um so I think that's very transformative like in a few years we

1:57might that might be like the main interface for working with you know any any data and it even writing code and stuff and jobs on the platform. Um and then the other thing it's like more sort of in the technical details but maybe like even even more interesting in some ways I I think is LTAP which is letting you have your Lakebase

2:20Postgres and your analytics um spark engine on a single copy of the data and you know it looks like native Postgres for the online apps it performs like that uh but you can also analyze the data like as it's you know as events are happening. We move so many like complicated systems and delays and stuff and like building an application um end

2:43to end. Um and it's it actually like will reduce overall the cost of like hosting these apps and let you understand much better what they're doing. Um and we're very excited by how well it's it's really performing and you know how we can basically turn it on for for everything pretty much. So >> I think LTAP is so exciting. I feel like it's almost like all the AI stuff is

3:05hiding how beautiful and crazy LTAP is, but just having one one system to do both analytics and real time and all the agentic things I think like it's it's really >> yeah and in general I think with AI and with with other technologies like you want basically it makes um you know data more useful right because you can make you know more decisions on it automatically can make better ones. So

3:29that means having like really timely data that you can then feed into these things is even more important. Like before if you got data very quickly uh but it was text or something like you need a human to look at it. So there's no point getting it instantly like just get it every you know every night or whatever and like take a look and see what's going on. But now you can have an agent look at it. So like yeah you could get it immediately and realize like this

3:53customer is unhappy or something and we need to do something um and actually take a good action about that. So I think a lot of the infrastructure will have to change and if you look at it with LTAP and with Zerobus as well you you can have basically a single table a single data store that supports like online postgress transactions supports

4:15like streaming records into it and supports analytics and um you know it's sort of a single copy of the data that does all those things and it's immediately visible to all those like use cases. Let's talk a bit about AI maybe and and what you see happening in the in the markets because I know you talked to a lot of CTO and CEO like we had the GPTs and now we have

4:37skills and harness. What do you see like what are they doing the all the enterprise all these companies what are they doing with that and what's the main challenges right now? >> So first of all I would say the the most successful use case is coding agents for sure. Um, and that's the one that you many people are adopting, but they're also worried about governance,

4:59security, and cost. So, that's the thing on top of a lot of people's minds. They they want to adopt it, but they're worried about how to do uh these things. And that's where like uh you know, products like Unity AI gateway uh come in to help you with those pieces. Um, I'm on the agents if you want to um give your developers the ability to use multiple agents together. Um uh but

5:22that's where they come in. Um now in other areas agents are picking up. So obviously we're excited about Genie and connecting agents to enterprise data uh with Genie ontology like you've got these agents that already understand you know where everything is and what it means and can just like correctly answer questions. Um and people c people people are building a lot of custom agents as

5:45well. And I think for custom agents, you know, we're realizing that um you need, you know, good evaluation and um good sort of context for them to do well. And I think enterprises are still learning how to best do that. I mean, just at Databricks, we're still learning the best way uh for different applications. But um I think it's becoming pretty clear that you can if you set up this

6:08feedback loop, you can get agents to a good level of quality. So I would say near-term it's definitely you know we have one killer use case coding agents how do I govern that and and you know make sure the result is still secure um and longer term it's like if that's working maybe I can build my own agent that does something interesting in my

6:32business and my setting how do I do that and um that's where you realize you need the evals and the agent design uh for for doing that >> is trust uh like a big issue nowadays like if you connect to Genie and want to ask a question about your business like you you get something but >> how do you so I mean one way that it's an issue for any agent is you asked it

6:55to do some work for you and it's like here's the answer you know it's 42 or whatever and then you wonder wait okay how did you get that you know what assumptions did you make what did you use so with Genie like we learned pretty early on that we want to provide good references, citations, we what it used and we wanted it shouldn't just say well I found this thing sitting there

7:18but it should actually say well this is actually the most widely used you know dashboard or table on this topic and the author is like you know this like principal engineer in that team or whatever like how do you make sure it's actually a trustworthy source once you have that basically it's a little bit of um user interface design issue in some Okay, once you have that, you can start

7:41um you know, showing users what you did and establishing that trust and helping them fix things if they're wrong. Um even if you look at Genie ontology, we have like these are like snippets of information that Genie learned as it was working on things and we tell you where it learned them from. So if it says like, hey, your definition of like leads in in your uh Salesforce is like this

8:06and you disagree, you can say, well, where did I get that from? and like shoot okay this you know this team is actually defining it the wrong way. I think that's what you need for trust and really it's the same for like coding and you know pretty much anything else if it did a lot of work you either have to like look at all the work or you have to be able to question it or like ask it hey what what else what sanity checks

8:28have you done to make sure this is correct >> and during the summit I remember we um uh Ali we mentioned the four C's the choice the fact that the platform can you we can serve any models almost all the LMS available on the get context you already mentioned it with the genie ontology and then we have cost and control that can be handled with unity AI gateway. So can you tell us more

8:51about this the last component? >> So Unity AI gateway is um is just this service that uh like proxies all your LLMs and all your tools like all your MCP servers um so that you can control like all the calls to them. You can log things and then analyze them. Um you can set guard rails on LLMs like don't send

9:14these type of inputs over. Um you can set uh you know policies on on tool calls and things that could be code or LLM based. Um, and also it's a natural place then to switch between models and to do routing. And we have at Databricks since we're a large consumer of like all the LLMs, we can basically like um, uh, let customers purchase tokens from any

9:38of them with DBUs. Uh, plus if you want you can use those DBUs for compute or for open-source models or like GPUs or like whatever you want. So it gives you a lot of flexibility as an enterprise instead of you're committing like a million dollars to like Google and then you realize hey actually their model isn't the best for what I want anymore and now I'm stuck with it. Um you just

10:00you know you can you can switch maybe on a per application basis and go back and forth. Um, so yeah, we think kind of like multicloud, but even more so for AI. Like you want a multi-AI strategy where there are these competing vendors that the interface to all the LLMs is like basically the same. You know, they they could all do any task. So you want

10:23to set up a market and like have them kind of pick the right ones for the right uh task. Um, and then on top of that, there's the security and the um uh cost management pieces. where tracking everything centrally uh really helps with that. I think it's kind of become, you know, it's kind of like having an API gateway in your company or or having a data catalog. It's like some kind of

10:47infrastructure that you want to have to use AI. Well, >> thank you so much. That was crystal clear. >> No worries. Yeah. And um yeah, thanks for having me on. In lake >> rise data se lake flow strings the flow iceberg and delta lake tables a glow unity catalog governs Lakebase oxql mosa and cha prolongs you and [music]

11:09co-host light the fire dream scale higher and higher [music] next drop