Why General-Purpose AI Will Wreck Your Database | Itamar Syn-Hershko, NeverBlink

Nowadays, everybody is using AI and AI is making software ship faster than ever. And depending on who you talk to, it is exciting, but nobody slows down to ask if your database can keep up with that. Now, there are more queries, more complexities, and not nearly enough expert to actually handle it in real time. So you may think the easiest solution throw some AIS throw some LLM at it and they will handle it and yes they do handle but what if not what if but they will hallucinate and when they hallucinate on production data there is no end do button that is exactly the gap never blink was built for it is the world's first AI database administrator and it just added support for click and today we have with us Sin Hushko, CEO and founder of Neblink to break down why you need your own kind of databased AI administrator, not some general purpose AI. First of all, Mar, it's great to have you on the show. >> Thank you. Great to be here. >> It's my pleasure. As I was trying to understand the problem and try to explain uh the problem and what you folks are trying to uh solve. So if I was wrong, please correct me and also try to explain what this company is all about. >> Yeah. So when you connect LMS to your database, you don't get a second chance. So effectively, if you're you connect LM to analyze and hopefully also fix any database issue when it actually happens, you can trigger a really bad domino effect and you don't want that. So you want something to watch over your databases at all times, let you know when something is about to happen or prevent things from happening and if something has happened to let you know how you can fix it um and be safe about doing that. So we built Neverblink or in the previous name that it was pulse uh way before AI was big and we had a lot of decision trees and things behind the scenes that can do a lot of smarts. So we effectively built a platform that can figure out do uh figure out issues do root cause analysis uh on read issues and find ones even before they become issues. And now uh we have AI to help us do things even better. And because we have that platform, we can be the honesty layer or the truth layer for any LLM wanting to know a database performance, database reliability status. >> Can you talk about why do companies actually need an AI database administrator? Can't they just rely on human experts because there are so many of those? >> Absolutely. So database experts are actually not that common anymore. It it has been a really big topic or or job u DBA right database administrators um for a long time. Um it's not that common anymore. Um and also a lot of companies don't really see a big need for that right so they would say okay we have a database it's mission critical sometimes we need to bring in a company to consult us on fixing issues or optimizing databases but it's not a big topic anymore nowadays um uh engineers are expected to fill in the gaps themselves especially in the AI era and there's a lot of managed services that offer a lot of those, you know, optimizations or removing the hurdle of running database on infrastructure yourself, but it's there's a side. So, there's a lot of things that you're actually not seeing as problems. Uh, which is the queries that you write, the table definitions that you do uh with open search, elastic search. We used to handle a lot of index mapping issues and so on. And sometimes you meet that those issues a little bit too late. Um so it's not a problem that you foresee having and then when you have it you are in some sort of a crisis mode that you want to fix at that point in time and that is what we are often solving as consultants and that's what we why we build a product that automates a lot of what we do as consultants. Can you also talk about why we just can't use general purpose LLMs like chat GPT or cloud or Gemini or there are so many out there uh to handle the database on it own why we need specialized you know AI >> yeah it's a good question uh we get that all the time as you would imagine um and the truth is those models they're really good they're becoming even better every point in time but there is two issues here one um and you can try this yourself. So we are discussing the click house uh new support that we've launched. Um if you go to chip, you go to claude um and you ask them a question about click house 80 90% of times you'll actually get it's not fullination but you'll get the LM to lie to you. It will not be as accurate and there's a simple answer to that. There is not enough information about click house uh out in the wild. There are documentation, the technology is still changing. Um, there are a lot of changes to the technology itself. Um, and we see that as cons expert consultants working with customers all the time. There is a lot of things that customers don't get right. Uh, sometimes because of knowledge gap, sometimes because they asked LLMs. Um, and and that's one. And the second, even if the LMS were good at that, you still need some layer of let's call it guardrails or called it uh honesty layer uh a few minutes ago. And effectively, you don't get a second chance. You want absolutely want to make sure that any decision made on your database to change it to optimize it to fix issues will be spoton. And that's why we built NeverBLink. Also what we have seen I am heavy user of AI myself and almost 90% of what we do uh this personal private work is through AI is that uh because of whole token cost uh and everything else the quality of you know AI is also kind of even you know some frontier models you know you opus or fable even you can see them working as good as open like pen you know so you cannot even and when it comes to database you And you don't have that luxury to so can you also talk about the the quality control the whole governance that is also required also when you talk about databases we can also talk about uh regulated industries we can also compliance so there is not that much room for error versus writing a code and testing it and doing something like that versus databases because uh apps can come and go database is what I would say is like your oil your solar energy that should not be shaken >> yeah Absolutely. So we deal with the operations side of the databases, right? And that side is super important and it's like I said it many many times it's something like an afterthought, right? It's so critical but people are just, you know, pushing ahead, pushing hard and a lot of times it's just easier to throw hardware or money at the problem. So just use larger instances and fix things like that. But there's a point where that doesn't work anymore and we are there to catch everything before that happens or and when that happens give you the the right solutions around that. Um we are working on features like um you know gating queries gating expensive queries the detecting that as that go on and um suggesting query optimizations and so on. So the idea is to make sure that the we are the depth layer of any knowledge gaps or missing knowledge um on the database layer. So you're you know any development team product team can move fast and not worry about the databases. It they didn't worry about it so much before anyway. the data platform is something that we see a lot of times as sort of you know ne negligent. So there's a lot of negligence around there. There's a lot of times where we get to a customer and the data platform is just not built to scale and it's you know being rigged as as we go. Um and nowadays with everything that's going on with AI there's a lot of shift left movement. So a lot of um product teams or a lot of companies uh software companies are expecting their engineers to do more and part of that is the database but you cannot expect someone to be just know everything. So they become sort of jack of all trades, master of none. And that's pretty much in the job description nowadays except you cannot really miss on the importance of of databases, the data platforms and that's what we are trying to solve here. >> Excellent. Uh thank you. Can you talk about uh what are some of the biggest headaches teams run into once uh database uh moves from kind of testing to production? Yeah. Um the scales the scales are just different, right? So any testing, any development will always be in smaller scale. Uh any query, anything will basically work there. Um also in production, you know, when you're small, uh it will work well and then at some point it's going to it's going to break. Um if you if you are an expert and you're looking at the database and you're reviewing queries, you're reviewing how the table is built, you're going to spot um issues early on. Um and then the idea is to find those things now in an automated way and make sure that doesn't hit production or we catch that before that happens. And then that is the biggest hurdle. Um, sometimes you know you grow into that in production. So everything works nice in production until it doesn't. And sometimes it's deploying something faulty from staging which is large but not large enough into production which is it's too large for what we're trying to do. >> You just added click house support. Uh can you talk about of course why click house and what does that actually change for teams running it? And then I would also like to know you launched it now so I will not ask about how people are using it production but I'm pretty sure you did some beta testing as well. Uh so I want to also know what their experience have been so far but let's talk about the first part u why click house and what does that actually change for teams running it. >> So why click house? Um it's an amazing technology. It's a really tiebreaker technology right. So it allows a lot of things to be done that you couldn't do before and the growth of Clickhouse um in you know in open source and also Click House Cloud the company behind Clickhouse that offers the cloud offerings um is just tremendous. I I don't remember a technology growing at that rate ever. I'm not that old but still um and given that I think click house is the great candidate for the first technology that we are we're supporting beyond open search and elastic search we are already private uh beta testing some additional technologies we'll discuss that in futures uh in future podcast I believe uh but for click house it's just an amazing technology being used very widely for mission critical ical um use cases and that technology is just not widely known enough not by people because the technology itself is still changing and definitely not not by LLMs because there's just not enough information uh out there about the technology. >> Now let's talk about what has been users feedback who have been testing it in beta. >> Yeah, so we've been working with a couple of customers. Some were already customers of ours but also have Click House. Um so we just let them use the platform and then some are dedicated customers. So customers who have been using um uh Click House on Click House cloud or self-managed Click House and they just needed a way to better understand what's going on and improve performance and most of the benefit for those customers were actually looking at the table hows are defined and how queries are sent. So it's uh you know ingestion into the click house and optimizing that or executing queries and making sure that the queries scan as little data as possible and run as fast as possible and that has been the main the main benefit. So effectively getting a lot of queries uh you know improve improving their performance um dropping the P99 and P50 to to in significantly to to you know one or or twodigit milliseconds in many cases um and then just making sure that everything works well for them even if they just don't know the platform. Well, >> can you also talk about uh of course people are running their shops with cloud you know of course with MCP and MCP's open source it has to totally change the game. Can you also talk about how well does uh uh you know your tool integrate with their solution how it works with what they already have in place. So when people think about MCP many times is connecting to a database and getting data from the database using an MCP which effectively translates things that you ask into something that you can do with the database kind of commands or queries that you can ask it. Uh we're going to release pretty soon an MCP that is aimed at operations people. So, SRRES which is site reliability engineers, platform engineers, DBAs and DBRRES. So, the people who are actually maintaining databases. So if if you are a database specialist, the MCP for of that NeverBLink is going to expose will let you get good good answers, grounded answers about your database performance, your database state and health and get those answers um in a very safe way and grounded way so you can trust it in doing your job. If you are an S sur you have wider capab wider responsibilities way beyond the database you can also reown that to give you answers on the database. But what we see a lot is that developers who are now assigned platform responsibilities and are not in any way experts in databases or data platforms in general. they can now, you know, move faster by relying on EverBLink to get them solid operation advice on managing and deploying and optimizing their the databases. >> Uh, can you also talk about your business model? >> It's pretty simple. Uh, NeverLink is a platform and the platform goes by what we think is a very modest and and affordable price and that's just a platform. You get you can connect your database. Obviously, we are stock to compliant. Uh we don't touch your data and an agent. Uh it's a terrible name for that, but it's still what we call it. You install an agent, which is just a metrics collector uh in your premises. It sends metrics to us. Um and then based we just optim uh we understand that and we provide those insight on the platform. On top of that, we provide developer support. So our experts are available to help your team build anything and optimize anything uh on top of Clickhouse and again open search and elastic search as well and more technologies coming soon. And then we also offer 247 enterprise support for click open search elastic search um that and that is on top of that. So you can think of it as a platform, platform plus consulting and enterprise tier uh which is platform and full support end to end 24/7 >> and uh can you also talk about the importance of partnerships for you folks? What kind of partnership you have with the ecosystem? >> Yeah, we work very closely with AWS. Uh we are under several programs of AWS uh isva and others and we work very closely with them. We're being brought into um places where you know our expertise is needed and also now uh where the platform can help with optimizing cloud usage and so on. as well as some um cloud researchers and uh so system integrators, other consulting companies uh that are needing are in need of of expert help and um a platform like NeverBLink that basically reduces friction and make sure their customers are um just moving faster and better with their data platform. >> It thank you so much uh for breaking this down for us and of course for the audience who are watching this. If you're running click house, open search or elastic search, please go to neverblink.ai and see it for yourself. Test it out. And once again, uh, thank you so much and I look forward to chat with you again. Thank you. >> Thank you for your time swap. And, uh, have a good time.