Let HeatWave Drive: The AutoPilot Advantage

In this episode, leFred and Scott are joined by Onur Korcerber to explore the many features of HeatWave AutoPilot. Learn how AutoPilot’s intelligent automation helps manage MySQL instances with ease, optimizes performance, and reduces operational costs. Onur shares practical insights and real-world examples showing how customers can streamline their database operations with HeatWave AutoPilot.

-------------------------------------------------------------

Episode Transcript:

00:00:00:00 - 00:00:31:20
Welcome to Inside MySQL: Sakila Speaks. A podcast dedicated to all things MySQL. We bring you the latest news from the MySQL team, MySQL project updates and insightful interviews with members of the MySQL community. Sit back and enjoy as your hosts bring you the latest updates on your favorite open source database. Let's get started!

00:00:31:22 - 00:01:03:00
Hello and welcome to Sakila Speaks, the podcast dedicated to MySQL. I am leFred and I'm Scott Stroz, joining us today is Onur Kocberber. Onur is currently a director of Development at Oracle, leading efforts on MySQL HeatWave, specifically working on the AutoPilot. Based in Oracle's Zurich office, Onur focuses in advanced research and development to improve cloud database performance through interpretable machine learning techniques.

00:01:03:02 - 00:01:24:16
He plays a key role in the ongoing growth of HeatWave, including work on new offering like the HeatWave Lakehouse and HeatWave GenAI service. Welcome, Onur. Thanks. Thanks leFred, thanks Scott. Great to be here. So Onur, can you tell us a bit about your journey? What led you to Oracle and specifically to the MySQL HeatWave team? All right.

00:01:24:16 - 00:01:53:10
So I, I was a grad student at EPFL Lausanne in Switzerland, and, I was doing research specific doing database, accelerators, both for, with hardware and software. And, at the time, I knew that Oracle Labs had a very exciting project about, building basically hardware, software, core design, database machines. And once I graduated, I knew that there were really good set of people.

00:01:53:10 - 00:02:21:18
And that's, how I joined. So I came to basically Zurich, to to the Oracle Labs branch. And then eventually, maybe fast forward ten years, we have, HeatWave database service, but, what we see includes MySQL and other things I will discuss today. That is fantastic. So, Onur, this entire season has been dedicated to, everything AI.

00:02:21:18 - 00:02:47:07
What AI offerings that HeatWave has and some of our listeners, I would guess maybe many of our listeners probably aren't too familiar with, HeatWave AutoPilot. Can you give us a high altitude overview of what AutoPilot is and, what problems that might be resolved? So the database systems today are all cloud databases, right? And, these are many services.

00:02:47:07 - 00:03:21:04
And the onus is on us, in terms of managing these systems. So the customers are expecting basically a full, full fledged, automated service with no, let's say rough edges. And that's where, AutoPilot, comes into play. And when we started the project, when, MySQL HeatWave was becoming a cloud service, we, also started the AutoPilot project, and, we basically targeted four different, let's say, problem domains.

00:03:21:04 - 00:03:53:04
So these are, setting up the system, data, basically loading the data or data management query execution and then failure handling. And, for each of these, categories, we basically looked at what, how we could, improve customer experience as well as customer performance. And at the same time, we put the machine learning, as one of our, basically main objectives because, this is a very old topic, right?

00:03:53:04 - 00:04:18:12
This is this is not a new topic like database management on automatic database, admins and DBAs and such. So that's why we took all the, academic research, plus the realities all today, which is the cloud services. And then, we looked at these four different pillars and then fast forward to today, we have like a double digit numbers in the AutoPilot suite.

00:04:18:14 - 00:04:55:12
Wonderful. And that's awesome. So and why then, this HeatWave AutoPilot is a game changer for users. Right. So, one of the things that we were seeing in the early days of our services that customers would sometimes put together, let's say, scripts or rules or let's say, some sort of, business practices, right? And in AutoPilot, we are taking all of those, especially what you're observing or what you're anticipating, right, that, the customers will have problems with.

00:04:55:16 - 00:05:18:07
And then we are offering them out-of-the-box ready to use for the for the customers. Some of those are fully automated, like, let's say, for or planned improvements. These are like these are happening completely transparent to the use it and some of the features that are a bit more about, the cost optimization of the service or performance optimizations are provided as an advisor.

00:05:18:08 - 00:05:43:03
So essentially we are constantly watching what the customer might, let's say, what would the cost of problems that the customers might have? And we are offering it out of the box included in the, in the service. And that is something, we see when we look at our competitors, we see that, some of the problems that we are solving are just seen as kind of still left as rough, rough edges.

00:05:43:05 - 00:06:02:08
And that's why it is really important. And at the core of it, we have a lot of machine learning models. These models are automatically up to...updated as we also update the version of the service. Therefore customers don't have to worry anything about, basically those, those, those problems that they are running into. Great.

00:06:02:08 - 00:06:31:10
Thank you. So, and when I follow what you just said, then, it seems that, these AutoPilot feature can save OCI customers some money, right? Right. So for certain cases, absolutely. For example, let's take auto provisioning. This is the feature that, the, made available almost at the same time when the, with the GA and, since our GA, this has been used, very actively.

00:06:31:10 - 00:06:54:02
And in this feature, for example, we say this is the number of nodes, that's, a customer should provision for accelerating their, analytical queries with HeatWave. And the great thing here is that, they don't have to overprovision their cluster or they don't, they don't need to under provision their cluster and then run into all sorts of possible issues.

00:06:54:04 - 00:07:13:07
So then one, one part of it is that they have the optimal cost, right? So they, they pay or they provision what they, what they should. And at the same time they also say, save time by just not having to, worry about it. And then similarly, for example, we have an auto load and unload feature.

00:07:13:07 - 00:07:40:05
So if you see there is some let's say there is going to be some benefit from from customer workload, we would automatically load or unload tables. And again, this would either give you a performance boost, which again translates into some sort of cost saving, or at the same time we would just, unload the unnecessary tables so that the customer wouldn't have to, let's say, increase their resource consumption, because they don't they don't have to.

00:07:40:07 - 00:08:15:12
And then we have a bunch of other like, similar features actually, that that will do. For example, there's auto compression that already gives you better price performance, but by default. Right. So that's definitely, every the most of the optimizations we do is translating into some sort of cost saving for the customers. That's awesome. I find that actually pretty, interesting that we offer ways to make sure the customer is basically streamlining their process, and then they're not overpaying for resources because some people might spin up a huge instance when they don't, in fact, need it.

00:08:15:14 - 00:08:39:07
So what are some features of AutoPilot that can help make storing and retrieving data a little bit more efficient? So I mean, let me give you an OLTP example. Of course auto indexing is is one of them. Right. So indexing, is definitely one of the holy grail problems in computer science, I would say. And we have a feature, that basically recommend secondary indexes.

00:08:39:07 - 00:09:04:23
So that's I see people ... people who are familiar with the MySQL know that how important indexes are. So we actually have an index advisor and that's, pretty effective. We see this today with customers as well. And that's just working really well. And having the right indexes is definitely making the, data retrieval, extremely efficient.

00:09:05:01 - 00:09:26:21
And if I were to give you an example from the analytical site, we, we have adaptive query execution. So we are basically over time, the improve the, the the query plan. Right. So this is also making, everything, a lot more efficient. And if I were to give maybe an example from the Lakehouse side.

00:09:26:21 - 00:09:57:14
So this is another, basically feature where we deal with semi-structured data. We do we, we automatically ingest, the unstructured files by understanding the, the, the schema. And, this way we can represent the unstructured data in the right format, which could translate into a better, let's say, space, usage guide so that you don't have to maybe pick a larger type than anticipated, than what the customer anticipated.

00:09:57:16 - 00:10:32:13
So and all these things, are they sometimes they look small, but these are the real problems because, especially when it comes to whether it's indexes or whether it is query plans or whether it is unstructured data, in all these instances, we are dealing with hundreds, if not thousands of either queries or tables and such. And and for a particular user, maybe dealing with 1 or 2 is easy, but dealing with thousands, I think every DBA would know or every user would know that it's, it's it's a tedious process with a lot of gotchas.

00:10:32:13 - 00:10:59:09
And corner cases will be basically take all these things into account in our AutoPilot suite. And then we update our, learnings and our optimizations as the versions go. But thank you. Yeah. Nice. Good answer. So what do you think are, the biggest misconception that the developers have about, machine learning driven, database optimization, right.

00:10:59:11 - 00:11:19:18
Because, yeah, there is the old DBA. Is that the they should know everything. And then the also thwy run the reports and sometimes people say, yeah, is it good or not I don't know. So do you know that, do you have an answer for this. Yes. So this is one of my favorite topics. Yeah.

00:11:19:19 - 00:11:52:18
So this is, something that we, we we have an internal discussion going on. Right. So I am also receiving a lot of requests from other teams or, people who are, and like, very, let's say, excited or ambitious about, applying like, machine learning to, to their domain, to their problems, like, one of the things that I keep seeing is that so there is, basically systems for ML and ML for systems that I, this is, I think, a very good, way of describing.

00:11:52:20 - 00:12:22:20
So we are at the end of the day, we are building computer systems and we should use ML for optimizing our computer systems. So and most of the time what I see is that like people who start, like basically trying to people who start trying to, apply ML, they put ML is a first object, whereas it should be actually not the first objective, it should be first the systems, how we build a system, a computer system.

00:12:22:22 - 00:12:46:22
And then we need to understand what is the hole, right, in our problem space that we can fill with machine learning. So most of the people who go and collect, let's say a data set and a the draw, let's say a regressor or a classifier on that data set. They say that it or it works well in the test, but it doesn't work in the in the real like a production like.

00:12:47:00 - 00:13:09:20
And to me this is the missing a systems inside. So we basically have to have a systems inside. So I will give you a very specific example. For HeatWave for when we are loading data into HeatWave, we can control InnoDB parallel thread and you know, parallel, like a thread count is a known lock that, anybody would know, let's say, how to tune.

00:13:09:22 - 00:13:32:01
But when it comes to the HeatWave load, it is, basically like the trigger behavior is changing, right? So basically we need to understand why the thread break here. The idea is changing. So that means that we have to collect the data in a way that we exercise the parts that HeatWave would exercise now. So if people were basically just saying, oh, like machine learning is going to solve everything for us, right?

00:13:32:01 - 00:13:54:09
Then we start with the data set. It will work. Basically, first the system inside, then the machine learning. You will be highly, highly effective. So that is why I think the second part is a while. Sure that what we put a lot of focus on is interpretability or explainability. So we try to fail our models first before the customers fail them.

00:13:54:10 - 00:14:23:06
alright and with that, we this is we just ship models, right. So this is, this is actually very, very important because otherwise it's some people might say, oh, you know, let's just fancy like a version of rule based tuning or it only works in certain cases. Right? So, basically to summarize, there is a lot of technical debt in machine learning models, but using the systems inside you can get good system engineer.

00:14:23:06 - 00:14:42:02
I think it's it's the real, secret sauce of shipping machine learning models that are effective at production. And then it's a long topic that after you ship it, you have to monitor them. You have to make sure that you are not regressing. There are not too heavy concepts and such. So yeah, what I would change that ML is machine learning is a tool.

00:14:42:04 - 00:15:12:07
And like any tool, it has its own drawbacks. And as long as we are aware of them, we can, these can provide really strong systems, that take advantage of ML. But again, systems first, ML second. That is my philosophy. I think that's actually a pretty good philosophy. So I know you might not be able to tell us everything, but are there any up any upcoming features that you're particularly excited about that you can actually talk about?

00:15:12:09 - 00:15:39:17
All right, so, well, I think this is not a surprise about generative AI. Right. So generative AI is, now the, the, the hottest, topic that, we are dealing with and, we, are working actively on generative AI based AutoPilot features, let's say. And one important difference there is that, when I say systems first, right, systems is about numerical data, right?

00:15:39:17 - 00:16:01:09
We deal with numbers that are coming in, let's say cache misses, buffer pool heat ratios, read write ratios right or performance and like all these numbers that they are essentially a time series that are just flowing in. And machine learning is like traditional machine learning is is very good at it or sometimes categorical, like, the data set, right?

00:16:01:11 - 00:16:23:22
When it happens, like should, it should have been this way or another. Right. Those are easy. But what is generative AI now bringing is, being able to deal with completely unstructured data. So what is unstructured data is text. For example the text. Then what text means means generating SQL code or text means dealing with logs, right.

00:16:23:22 - 00:16:53:15
Or log files or or automatically, thinking into, let's say like, your own, let's say diagnosis. That goes to data insight. So those are the areas, let's say that, that we are working on. Excellent. So because you're talking about, AI, GenAI, the resources is something, we hear, more and more in the database, area.

00:16:53:16 - 00:17:21:12
So, because HeatWave AutoPilot brings us, already a lot of, intelligence, automation but, what about the the natural language to a SQL. So the NL to SQL, do you think, this will be also something that, will come, for us and, and do you see it as a serious productivity tool for the analyst and the developers?

00:17:21:14 - 00:17:52:18
Are there still, orders, to making it available and, enough, for a production use? Right. That's a very good question. Okay. This is definitely a very hot topic for for everyone in the industry, I believe. And, so, yeah. So what is really happening in the NL to SQL domain is, as the large language models are getting larger, we are seeing a very big improvement in the accuracy of these tools.

00:17:52:20 - 00:18:36:02
And of course, what I mean, my accuracy of the known benchmarks. Right. And also what is really... Another interesting trend that is happening is that, you know, you look at traditional, like benchmarks, like for OLTP, OLAP, you know, TPC-C, TPC-H, TPD-DS right? They've been around for a very long time. What is interesting about, this NL to SQL benchmarks is that the more that they're out, it's, it takes maybe a year or less than a year to, to to get really good scores, you know, like, people are conquering is benchmarks and they're very fast, pace so it is true that it was not ready, but now we are seeing

00:18:36:02 - 00:18:58:06
signs that it is actually, they are getting very, very good at it. And and of course, it really depends on your, your complexity, how many SQL constructs you have. Right. And how complex, it could become because you certain SQL is just pages. Right. So that that means you need an assistant. But certain SQL is just, you know, you just want to learn something about your database.

00:18:58:08 - 00:19:17:02
Maybe you're a business executive or maybe you're a data analyst who is not very well versed in SQL. I can tell good for for for those type of use cases where you're a business analyst or whether you're a data analyst, I think that the tools are definitely there. So more complicated queries where even humans don't write in one go, right?

00:19:17:02 - 00:19:41:15
Like its an NL to query it it's pages right there. It's a great system, but it is performing maybe similar to the coding assistance that that we have todat. But yeah, it's it's definitely in a corner like, it like is is basically something that everybody is looking at actively and so, so we are so that's, I think you will hear,

00:19:41:15 - 00:20:04:08
Hopefully, some more cool news about it, soon.. That's awesome. That's actually something I'm really interested in. Just because, like you said, allowing people to query data without actually having the SQL knowledge is, is kind of intriguing. So one last question. As you, as would the three of us should know. You know, here at Oracle, we tend to eat our own dog food.

00:20:04:10 - 00:20:33:12
And it helped improve MySQL in some areas like high availability where we tuned group replication. Do you have a similar experience with the AI related tools that you could, talk about? Oh, yes. That's a very good, question. And this is an active topic that is, basically, happening right now, within the MySQL Org. So, yeah, I'm working on two different projects.

00:20:33:17 - 00:20:59:02
One is, we are using generative AI, our own generative AI service. I think they've generative AI service to generate, HeatWave release notes. So if you go to mysql.com today and if you look for HeatWave release nodes, you will see that there is actually banner up there that says these nodes are generated with the assistant off assistance of, HeatWave generative AI service.

00:20:59:04 - 00:21:20:12
And this is a system that we built in, again, purely running on our, software. And, it's working. Our technical writers love it. And we have actually working on several other improvements that we we are trying to write more, with that. And it is going to come. So there's something of the, quote unquote drinking our own champagne, right.

00:21:20:14 - 00:21:50:14
And, with that also reflected back right to the, to our generative AI team. So the AutoPilot team initially basically said, okay, like, these are the things that you should improve because so if you're running into these problems, our customers will also run into these problems. Right. So that's that. Is there one one one good. Also way of let's say, improving our product and another one, it's, something we call Ask MySQL Expert in short Ask ME.

00:21:50:16 - 00:22:17:17
So we are we is to build a general, question and answer machine that is able to, aanswer questions like the how to questions or the troubleshooting questions internally, like within the within the MySQL org and some of it's actually we, we demoed in some of the keynote speeches that that we gave and the also recently released a version of this as a sample app to our customers.

00:22:17:19 - 00:22:40:05
So basically, one part is that it's a question and answer machine that we are using internally. And, we, we got really good feedback. There are, certain cases where, especially junior, engineers, they learn a lot because they need to onboard, faster, than let's say, you know, basically compared to the compared to the past.

00:22:40:05 - 00:23:08:11
So that is why I think because there's a lot of knowledge to, to carry up to the you to use this tool. And at the same time we release this version to our cust... to our customers. But this one is something that they could bring their own data to build their own. Asked me let's it right? Of course, in this case they could call the tool something different because the case is something they could just, extend so they can just open their own data and, and build this, question and answer, let's say, machine.

00:23:08:13 - 00:23:33:19
And IT will, of course, we are evolving these tools forward so that eventually some of it could be also be part of our, our service. So and, yeah, these are actually two specific examples where we are using generative AI actively. There is a lot more, the only I think, you know, limitation in this case is like, is for us to catch up with this technology, right?

00:23:33:19 - 00:24:00:14
Because the Geratvie AI space is moving really fast and identifying what is working and what is not working and what is useful today. Our what is just forward looking. It is maybe let's say 10% or of of of of my time, like just so that we are not always also working on, very big, let's say, blue sky ideas as opposed to kind of making people's lives easier today.

00:24:00:16 - 00:24:28:03
Thank you very much. So yeah, I saw during the MySQL summit the example of the, of the app, with, the knowledge base, but for example, for the HeatWave release note that I wasn't even aware of. And, after this recording, I will watch it immediately, just for fun. So, to wrap up, I want to I have something in mind that I want to ask you, if you agree with, but, with the MySQL HeatWave service.

00:24:28:03 - 00:24:52:10
Right. We have AI that, so you can build your AI application or, AI anything with the vector search. You you have, you know, HeatWave that will help for you doing AI, but it seems with the AutoPilot and stuff we have also AI that helps, MySQL HeatWave users to improve your experience. Right.

00:24:52:12 - 00:25:28:18
So that's a very good distinction. And I have to admit that sometimes I am also not, let's say making, like, drawing that, like, you're absolutely right. So some of these applications we built are for customers and they could just go and extend it, or they can take inspiration from that and build something else. Right. So that's we're demonstrating our own, technology that, when it comes to AutoPilot, it is essentially another way of, let's say distributing this application, but it's all under control and it is out of our service that it's building to our service.

00:25:28:20 - 00:25:50:19
And you're absolutely right. In both cases, we are using similar, ideas or similar technologies. And, one of them, again, we are giving it to the people so that they could go and extend it the way they like and the others is is our under control. We see how people are using it and we are improving it as as we go along.

00:25:50:19 - 00:26:11:04
That's why there's always a new, let's say, auto feature. I mean, coming out and then sometimes the lines are actually pretty blurry, like, okay, I don't want to make things complicated, but one thing is that we are also working on some intersection where let's say that you have an application that, let's say, uses AutoPilot in a way that let's say a user would interact, but.

00:26:11:04 - 00:26:33:21
Right. So those these are interesting boundaries that we are always, looking. But from the product communication or maybe from the road network perspective, we, we don't really talk too much about this kind of stuff because it's just, maybe it's a little bit more like an intellectual exercise for us to see the limits of our technologies. Thank you very much, Onur, for, taking the time to talk with us.

00:26:33:23 - 00:26:53:10
I thank you, thanks for having me on this. Thank you. Onur. That's a wrap on this episode of inside my exclusive Killer Speaks. Thanks for hanging out with us. If you enjoyed listening, please click subscribe to get all the latest episodes. We would also love your reviews and ratings on your podcast app. Be sure to join us for the next episode of Inside Mysql:

00:26:53:10 - 00:27:05:18
Sakila Speaks.