The real-time database to build your AI future (sponsored by Aerospike) (Aerospike )

Hello. And good afternoon, I'm Linley Hensarling. And with me, I have Amit Shah from Amazon AWS and Shay Peretz um from Risk I and A, a rule um ha had wan um from um free will.

So we've got a, a great talk today. I'm gonna talk a little bit about the pressures of A IML and what we have to do to, to really apply it in the real world. So there's, there's a extreme pressure to exploit this new technology, but it has a countervailing issue of cost that comes with it and the need for ever more data to be applied.

And we've got a number of people here with us today that struggle with this and deal with it and solve this problem every day. And we'll, we'll get into that as we, we go into the presentation. But a I can be a great cost savings in terms of what you can do with it, how you can make your business more efficient, but it comes with expense in the real world in terms of the computing power, the storage you need to use to use it effectively.

So it's interesting to look at some of the stats on how much computing consumes of overall electrical consumption. It's gone from 1 to 2% in 2018. And this is like in a very few years and then in 2 2030 it'll be in the multiple digits 8 to 21%.

Training in a I model um uses a ton of carbon and then the industry in general is bigger than the airline industry in its energy consumption. But we save a lot from applying technology to the problems of the world. And I say it that broadly that a I is going to solve many problems including energy consumption. But we have to be conscious at the same time of spending that and a lot of that is dealing with the massive amounts of data that have to be applied.

Ok? And it's massive amounts in generating the ll ms but also in the operational data that has to be affected by the application of the ll ms. And so we're constantly as enterprises ingesting data all the time because none of this is of great value unless we can apply it in near real time.

One of my jokes, I've been had a long career in tech. I feel like I've been chasing the present forever. Ok. What we're trying to do is have context for what's happening right now and be able to apply it right now. Ok. So being able to ingest to build models to apply those models in real time to get answers as things happen. But with the context that's required to give an accurate answer, given the state of the world, the state of your business.

So you're gonna hear from two of our presenters today about how they use aerospike as a feature store and why they chose to use it. And it's a question of efficiency and being able to apply more data in a decision process in a time constrained sl a and then offline stores where you're in constantly ingesting new data both for training but ingesting data that's happening now that you're going to apply those features against to make decisions.

And the closer you get to synchronizing all of that, the better outcomes you have. And there's always the metadata about the data you have that you have to keep track of. So podcast is a company that uses us for for these and you know, they have 10 billion records store data of eight terabytes and applying that for the historic contextual stuff and then get lookups in one millisecond so that they're never very far behind what the customer is doing or the data they have on the customer or the data they have on the context.

We also have an ecommerce provider that is in japan, who's a fashion provider and they have to provide the context in real time, but they have to be able to know that taylor swift changed the color of sweater she wears so that they put that up in the basket rather than, rather than the one she wore last week.

All of this takes a complete data platform. So we have an amazing real time database, but that's not the complete story because we have to provide the ability to fit that database into an overall data pipeline with multiple points of ingestion. Every one of our customers has many, many different um uh components that are coming in. And then you also have to be able to shovel that data to where it needs to go for learning and then be able to apply it.

We've recently added a graph capability with all those same characteristics. We're also working on vector and we'll do vector a little differently because we care about this time bound sl a and the ability to apply whatever data type you're using to drive decision.

So think r a ok. And being able to apply that in real time against the database of vectors that is commensurate with the context you have to apply to the generalized ll ms. So we also care about efficiency and have exploited graviton. You'll hear a little bit more about that from, from a me.

Um and we're architected for efficiency. Now, the big thing that we're focused on is being able to take advantage of whatever instance types, whatever types of, of networking that the cloud providers are giving to us and exploit that but we do it with the ability to scale out to scale up with specific technology that we have and we support geo distributed transactions.

So think synchronous active, active. So we can do transactions west coast to east coast on amazon. We do this today for payment providers where that you can lose a whole region, you can lose a whole data center and it keeps working and no data is lost. And we have massive parallelism and this is at the heart of our ability to support the tremendous throughput that you'll hear about from shy and usual.

So with that, I'm gonna ask a me to come up.

Hello. Uh can you guys hear me? Awesome. Hey, uh um my name is Amit Sha. I lead the uh product management for AWS cc two products and um about three years back uh when I got, uh when I joined AWS, uh I was working on storage and memory technologies given my background.

And one of the first things that I was asked to do was figure out uh what are the most stringent use cases for memory and storage uh who really stresses them and then figure out a way to optimize them and guess what the people that I worked with are sitting here in the room, shrini lindley. Uh I, I worked with them um over, over the last three years because they really uh like latency is super, super critical for them.

So I'm going to talk a little bit about what that journey has been um aerospike in AWS and, and uh how we have partnered together uh to bring uh real time use cases for our customers.

So, so based on what uh len le already talked about on, on the various customer needs and the, and the fact that aerospike wants to embrace machine learning and, and generate a a I aws wasn't, was an obvious choice when it comes to performance. Uh aerospike is a real time database who cares about very high throughput at very low latency. On top of it, you add the fact that you want to add machine learning capabilities, which is integrated through the um uh store offering as well as vector db.

They wanted a high performance cloud uh to meet their customer requirements, dealing with millions of transactions per second or sometimes iii i in fact, have worked with a respect, customers who are dealing with billions of transactions per second at a very uh at a, at a millisecond or less amount of latency is very uh is core to their, their business and to their customers. And that is where a ws really comes in with the variety of offerings that we have.

Um so from a performance perspective, how do we deliver all this transaction that can scale up based on demand and can can meet the customer sl a. So that, that is important but not only performance uh from a customer perspective, they also want to make sure that the cost is optimized ultimately, uh aerospike, uh the the idea of aerospike is that they want to deliver highest performance at the best price performance advantage.

So, so that that was core to uh to how you want to enable aerospike on AWS um delivering leadership performance, delivering leadership cost profile. Apart from that, uh the customers um that use aerospike on AWS, they deal with tons of data, meaning you're talking about petabytes of data uh get and the scale of transactions that I already talked about.

So when you, when you have so much amount of data to be to be crunched, not only you need good assessors and a significant amount of memory and storage, you also need network and and high like big high throughput network, low latency network so that you can deal with all these transactions. And that is where AWS system also comes, brings its own advantages.

Uh the last thing um aerospike wanted to make sure that the cloud provider enables their solution globally and can be secure and reliable, meaning their customers are um they, they have variety of use cases across financials, uh manufacturing real time and all of these customers, they have footprint all across the globe.

So for them, they need to have ability to scale across across the globe to meet their end user reach and that that was again important for them. So as far as AWS is considered, we offer aerospike and their end customers, many many choices choices in terms of uh what we can deliver uh uh in uh computing capabilities.

Uh here, when I, when I say computing capabilities, I'm talking about selection of processors, the amount of memory capacity that you can have the memory ratio so they can optimize for cost storage network choices. We are talking about operating system, payment options, payment plans so that there are many, many choices integrated into all our offerings.

In fact, we have one of the richest portfolio of products within cloud with 750 more than 750 instances available on cloud today. Apart from that, uh when we have a modern two instance which is built with AWS nitro system, then it does support highest possible compute performance.

What I mean by that is that a two instance with AWS nito system will deliver up to 15% more compute performance compared to a comparable instance from a another cloud provider which does not have a nitro. So, so there is definitely an advantage that you get in terms of price performance and performance benefits.

So, so so as far as the journey of aerospike is considered considering that they are a real uh real time database use case the two big things um aerospike cares about is high throughput at a given latency and while meeting an sla of about 200 microseconds.

So uh so given, given the fact that latency is so important and latency variability is so important. they wanted to make sure that the storage aspects of our offering, it says that can meet their sl a as well as optimize the tc for their customer.

Um the the journey of aerospike has started with AWS using the core database that is still predominant uh that is still the predominant usage of aerospike on AWS uh for our mutual customers that said over a period of time, as len le mentioned, they have enabled aerospike cloud graph and vector db on a w as well. So the offerings have just gotten richer where more options are provided to our customers.

In terms of the instances itself, i four ie two instance i four, i and i three e are the work horses. they, they are able to process millions of transactions per second and while providing very low latency. So, so these are the, so the real real popular instances that run aerospike use cases apart from that uh when we were developing nine to ssd.

So a w 92 d is our own custom built uh storage solution. And i remember um uh when i, when i joined AWS and, and we started looking into, ok, how do we make sure that a nitro or storage solution are really going to meet the most stringent requirements that are um that are available uh or being run on AWS cloud. And one of the use cases that really came uh came across was the aerospikes real time database use case.

So here we wanted to make sure that the latency and the latency variability is as tight as possible. And for that aerospike had created something called a ct aerospike certification tool test to test the storage latency.

So we, we partnered with aerospike and we optimize our night to ssd specifically for API use case where the new generation nsd based instances offer up to four times ac t compared to previous generation i three instance or up to two times a per drive compared to previous generation i three instance i three e instance.

So generally, a lot of work had gone in to really make sure that we can optimize these use cases for sike. Not only that we were able to reduce our, our latency significantly says that aerospike was able to see 70% more performance for a given latency, which ended up resulting in 25% improvement for customer t which was a huge benefit for the end customer.

So, so that was that was great as well. And our partnership with aerospike does not really end here. We we worked with, with aerospike on enabling graviton which delivers the highest computer performance in cloud as well as delivers the best price performance.

So, so a lot of work has been gone in to enable those solutions. So if you have, if you, if you do want to benefit from highest performance, you have sustainability, sustainability goals, then aerospike with graviton would be, would be an awesome solution.

And apart from that, as aerospike continues to engage with machine learning and, and embracing that technology, we would be partnering with them to make sure that all the future offering are integral part of how aerospike is optimized on AWS.

Given that um i would like to invite uh shi from uh risk. Risk fight. Sorry, shy. Here you go.

Thank you. Uh hi, everybody. Can you hear me? Yeah, so uh nice to meet you. My name is Shai and uh I actually fill in for a guy who's supposed to be here today and he prepared the session and everything though. Uh because of the situation in Israel, he couldn't make it and asked me to fill in for him.

So i hope that i will do justice to his short session. Uh I'll talk a little bit about Risk. I, what we do. Uh uh it works so Risk. I is a company that exists for the last 10 years or so. A little bit more than 10 years uh and went public two years ago.

Uh we have about 750 employees. Uh most of them are in Tel Aviv office, but uh the rest are spread around the world. We have a big office. The headquarter is actually here in New York, here in New York. Uh and uh we're spread about 180 countries where we provide services to customers.

We deal with a lot, a lot of uh transactions per day. Uh and a lot of gmv that we actually provide services to which i'll go into in a second. Uh and we have very, very good retention rate. And i think part of it is because the architecture and because the, the business model that we have developed over the years, uh you can see the list of some of our customers.

Uh we s we mainly provide services to very, very big merchant around the world uh that have a lot of traffic and it's uh nice to talk about it in cyber monday, which is prob probably the highest. Well, there's a competition between last friday and today, which one is the highest traffic day for those merchants and obviously for us.

So the fact that i can stand here and not in the war room or something is, is good. It means that the system is reliable and stable and so on. Uh so what we do, we have a platform of services uh that helps all those merchants fight fraud.

We collect all the, all the customer data which is related to the orders. We get all the information from all the customers and uh we run our magic around it in order to say for each one of the transaction, if the transaction is fraudulent or not if it is, uh we will say so to the merchant and they will decline the transaction. And we also take the liability, which means that if we decline the transact transaction, sorry, the opposite.

If we, if we say the transaction is good, it's ok and it's eventually uh that transaction and it's fraud, we will pay for it. Uh in order to be able to support this level of liability and responsibility, we have to be very, very accurate and to trust our system.

Uh so this is why we had to architect it in a way that we can trust it so much. Um this decline, approve of uh deals is what we call chargeable guarantee, which is our main product we also have in the platform. That's not for me.

Uh we have some other um products that complete it, complete the platform which deals with uh preventing account takeovers, uh protecting abuse of policies in customers. For example, if they want to limit the the number of uh items that they sell in a certain geography or if, if they want to uh fight uh people that ship back, empty boxes for example, or fill them with sand and send them back.

Uh all kinds of things like that happen. And we know a lot about the identity of the customers and we are able to identify what's the probability that this will happen. And when it happens, we signal to the customers and they can fight it and can prevent this, this form of fraud as well.

In order to do that, obviously, we co we collect huge amounts of data from all the customers. We link all this data together. And i think that one of the biggest effect that we have is what we call the network effect, which enable us to uh try and look in all the data from all customers and identify if there's someone which is believed to be a fraudster or abuser of a policy.

So we have all those domain. This is a very schematic uh figure of uh how it works, what the flow and what the problems which that we're trying to solve. It's just a very little flow from all what we do. Uh but this is the one where we chose to uh adopt aerospike as a solution uh to begin with.

We are fairly new partners with aerospike for the last two years. And uh and i think it brings a lot of value as you will see and, and hopefully we'll be able to use it in other use cases as well.

So domain services are all the data collection that we do in the integration with all customers, bring the data in all this data goes into our platform. And we need to be able to uh uh provide this uh recommendation of decline or proof in real time.

Uh why real time? Because we have uh some customers like grubhub or doordash, which needs their uh answer right now. Otherwise it will not be relevant. Some other customers don't care so much so it can come a little bit later so it could be like online or a little bit offline.

Uh we do all this in an order flow. Uh we created what we call ofe which is order flow event, which is signaling the status of the order on its way to uh to production what we need to do with it. And uh we needed to create some kind of a source of data that would be one of for everything that we do with the data later on.

This is this goes both to the live and online systems like the real time decisions. And it goes to the model training, a lot of what we do relies on model training a I models and obviously it should go to the billing and should speak the sa me language because it should be very, very accurate you build for exactly the transaction that we processed and everything.

And the rest goes to the b i systems to the da data lake and then the b i system. So we can use it for uh financial services, uh business analyst analytics and so on. Uh if i take a look, a little bit of a deeper look into the architecture, this is where we uh actually about two or three years ago, we came to the conclusion that the uh previous technologies and architecture that we use will not scale and will not bring it to where we need it to be.

And we started to look for solutions and uh and happily found aerospike as a good solution for those situations. So again, we have the service domains uh there in the left collecting all the data, then it goes via kafka into our of events processor. And this is kept as practically as raw data to the first instance of aerospike, which is really, really fast and reliable and, and really provide a very good solution for that.

From there, we have all the services that you, we, we use in order to enrich the data aggregates the old journey and everything that we need to do with the data in order to process it and prepare it for the next steps, then it goes and being stored in another instance of aerospike from there.

We can provide the services of real time data decisions and ml model training. And this is like a loop closed loop that goes together and provide uh better accuracy over time. And when we talk about like a I models, uh we train a lot of models, we have uh high segmentation because we have a very uh spread uh amount of customers from all uh different geographies and from all the different spaces.

Uh so you need a solution for each one and if and even for the same.