Rebecca Jozwiak: Ladies and gentlemen, hello and welcome to Hot Technologies of 2016. Today we’re, “Exposing Differentiation: A New Era of Scalable Infrastructure Arrives.” I’m stepping in for Eric Kavanagh today. I’m Rebecca Jozwiak, your humble host from the board group while Eric is off in Jamaica. Good for him.
So, like it has been for decades, this year is hot, although arguably technology is moving at a pace that’s outpacing Moore's law, and what are organizations doing to keep up? They’re looking for whatever is fast, and scale, I would argue, is probably one of the most important things when we think about databases. And of course we have the options to the usual relational, now we have our NoSQL, we have our column store, we have our graph databases, our RDF databases, but really, what businesses are looking for is scale, is parallelism and is fast.
Now, traditional architectures were kind of based on that relational model. But if you look at most web business that have sprung up in the last three, five, ten years, that’s not the models they’re using for their infrastructure. They’re using a different, a parallel architecture, they’re scaling and they’re fast, and that’s kind of what a lot of people are turning to, today.
Our lineup, we have Dez Blanchfield, he’s a scientist from the Bloor Group. We have Doctor Robin Bloor, our chief analyst at the Bloor Group, and we have Brian Bulkowski, CTO and founder at Aerospike. So guys with that, I’m going to turn it over to Dez.
Dez Blanchfield: Thank you, and thanks for having me here. I’m going to try and set the scene for how we kind of very quickly got to where we are, and we’re going to dive into a lot more of the technical detail as we go through today's topics. I’m just going to get control of the screen here.
So bigger, better and faster. When I think about where we’re at, the image that keeps coming to mind for me personally, is this very image that I’ve got on my title slide, which is the expansion of the universe. We’ve had technology develop and grow for decades now, in fact from the late fifties when the mainframe became a real thing. Technology has continued to grow in many cases at a worse or greater than a linear curve, depending on which part of the curve you’re on, as far as the software or hardware goes.
The scale has gotten bigger and bigger, and faster and faster, as far as what we’re trying to deliver, and smaller and smaller at the manufacturing and semiconductor level. And in the middle there’s software and the applications and the systems that underpin that software, and they tend to get smaller and smaller in nature, and we’ve seen things like containerized applications and micro servers, it’s become a thing again. We did that in the past, decades before, but as a result of going smaller and smaller there, we’re getting bigger and bigger at the scale of which we can now run things, such as applications and particular databases, and the logic of those databases.
I have this view where we’ve scaled very horizontally, in essentially the X axis; we’ve scaled vertically in the Y axis. We’re at the point now where we need to go somewhere different, and in my mind that’s sort of mentally envisaged as a Z axis, and that is that we’ve got to go deep into the technology and look at how we can do things differently to what we’ve done so far, to get that additional piece of speed. So I do visualize this whole expansion of the universe, where we’ve had an explosion take place, and some technologies exist, and this better linear growth and demand. We’ve had to find different ways to get that bigger, better, faster result.
Just to quickly cover kind of where we’re at now in a couple of hardware environments. We’ve seen the falling costs of a gigabyte of disk space bring about a couple of fairly large transitions and technology, and approaches to the bigger, better and faster scale issue. These are two separate graphs that cover roughly a decade apiece, just over a decade each of the falling price of one gigabyte of hard disk space.
It’s a classic J curve or a hockey stick as we often refer to them, in that some time ago you could spend literally hundreds of thousands of dollars to buy a gigabyte of disk space, not quite two decades ago, whereas today it’s become dollars and eventually I’m sure it will end up, what we term the race to zero, it will become cents. That brought about an interesting change in the type of things that businesses could do. And I refer to that as a disruption through data or big data in particular, and by that, what I mean is that we saw technologies, like how to become a thing where we could scale very horizontally in storage, and the type of compute we can apply to that storage, and how it opens an interesting technology because it allows us to do very large, redundant parallel storage at the fastest level, and Hadoop parts in itself, natively being able to copy data in a write once read many times format, and just scale the thing out at a near linear grade.
And it’s all companies like this come about true to disruption using big data. We’ve got companies like Uber who’re the world’s largest taxi company. They don't actually own any taxis, and it’s a long list here. Airbnb is the largest accommodation provider, actually has no real estate. One of my favorites is Facebook, for example in this list, where they don't actually create the content, we create it for them, but they’re actually the largest media owner on the planet. We’ve got interesting ones like the fastest growing banks, actually have no money. These are peer-to-peer lending platforms and banks, and there’s one in Australia in particular that’s growing fame here called SocietyOne. And some of the major banks who do have to have cash are investing in that particular peer-to-peer bank. And we go through this list even down to Netflix; they don't actually own any cinemas and yet they’re effectively the largest cinema house on the planet.
So they got to where they were, in my mind, through the application of smart technologies at the data level, because we could do bigger and broader storage at lower costs because of the fallen price of a gigabyte of hard drive space, and we could apply some intelligent compute and distribute a computing model over that. These companies had the ability to create a competitive advantage and disrupt as a result of that falling costs of disk space.
We’ve seen a similar thing happen in the cost of memory. A couple of decades ago, if you had six million dollars lying around, you could buy a gigabyte of RAM, and we’ve had a very similar J curve or hockey stick, take place in the reduction of the costs or the fallen price of RAM. And that’s brought about some interesting things, and in my mind, one of the greatest disruptions in that space is the amount of memory that’s being able to be built into devices, like mobile devices, like phones and tablets, and even laptops. Computers these days, the amount of memory that goes into an average laptop it’s quite ridiculous in some cases. In some cases, my current laptop has more memory than some of the servers they used to use not so long ago.
This has brought about significant change in its own right, in a similar way that [inaudible] a RAM has in my mind, it allowed us to scale and scale rapidly. And now we’ve had the emergence of a technology that we call flash, and this is a technology that originally stems from something that’s sat on hardware in the form of an EEPROM, a little chip that was designed to be able to be available, and write to, and then just when the power went off it would keep whatever you wrote to that chip as persistent storage. It was slow, it was clunky and in those days, I think it was about 1980–1981 it sort of became a thing. By 1984, Toshiba who I believe invented the technology, made it a commercial thing that we could use.
But before long, people figured out they could actually take a combination of the components that were used to create this concept of an EEPROM, a read-only memory, once it was erased it and written to it, and they could actually write to it on a regular basis, and use it a bit more like disk space, and a bit more like RAM. Over time, it developed. Now this flash storage technology has been a mergence between traditional disk storage, whether it’s a spinning disk or in some cases a hybrid disk of memory, and RAM. And the key thing is the system between because you can read and write to it, and then turn the power off, and it will retain what you’ve written to it. So a disk space, obviously you write to it, you turn the power off, and the spinning spindle and the heavily modified [inaudible], for want of a better description, keeps the zeros and ones you’ve written to it.
In the random access memory space, you write something to memory in RAM, you turn the computer off and everything gets wiped because there’s no more electrons to keep it charged and hold the information you wrote to it. Plus it’s in the middle and it’s extremely fast, faster than disk, a tad slower than RAM. But you can write to it, and read from it, and when you turn the power off, it will persist. This has brought about some amazing technologies and particularly we’ve developed mobile devices and laptops that are really, really fast, and able to do lots of things, and now it’s moved into the infrastructure space around storage and compute, and that’s brought about significant changes in what we can deliver at scale. This is kind of where I believe the Z axis in my mind is coming about now.
It’s almost just in time in many ways, because we’ve seen a disruption now through what I refer to as demand, and that is that consumers have, irrespective of what’s happening in the infrastructure and the technology space, and the ability to drive faster and more rapid compute, and performance at the infrastructure level, consumers are demanding this disruption in the form of what’s referred to now, the celebrity experience. Everybody wants every system, every app, every website to know who they are and what they like, and to be able to give them a personalized one-on-one experience. It’s not good enough anymore just to go to a website where I buy cinema tickets. I want it to know what I’ve bought before, why I bought it, and potentially what people just like me bought and recommend things.
Invariably, we’re seeing what I refer to is a side order of social, and that is that I want the celebrity experience, but I also want to socialize that idea, I want to share it with all my friends and tell them what I’m doing, and I also want to know what my friends are doing. And this is a result of an explosive demand for additional compute and storage, and rapid turnaround of things. We’ve seen the Fitbit generation, what I call always-on tracking. Everything I do gets tracked, and logged, and captured somewhere. We’ve seen real-time everything: banking, bidding, recommendation engines, having to be able to cope with real-time things I’m personally doing as the consumer.
And then we see a very big impact, like the security risks around cyber security. It used to be that we had individual hackers, then we had criminal gangs apply themselves to it, now we have entire nations going to war over the internet, which is a real thing and actually happens. Pay attention to that, sit up and have a look at it, because there’s a real impact to that, and some of our pre-show banter was around discussing the risk of having your own computer, or at least your network, penetrated.
We’ve seen this concept of entity extraction. Entity extraction is when we’ve got to find things of interest inside very large data sets and particularly around fraud, and illegal, and hacker-type activity. But more often than not, we’ll see that entity extraction is becoming a focus point for good things, and things that are of value to us, as opposed to looking for things that are attacking us.
We’ve also seen an explosion, what’s referred to as geospatial data. This is data that actually knows where it originated from, or where other data like it is from. You can imagine you’re standing in the street and you want to find the nearest parking station, or the nearest restaurant, applications that can apply geospatial compute and data, computing to data, that knows where it is in space, is very important because you need to be able to know where other objects and entities are, and do that quickly.
We’ve seen permanently connected mobile. Even when we go to sleep at night, our mobiles are still ticking away, updating our emails, checking our calendars, looking at what the weather is and figuring out what whether what we’d like for breakfast is going to be available. There’s a lot of noise happening there, and that’s created a massive impact on what we need to do at the back end, and how fast we do it.
Overall, the sheer scale and impact of what’s being referred to as the Internet of Things, or more often than not, the machine-to-machine connectivity, where devices are talking to devices and that goes all the way up to engines strapped to the side of airplanes telling the airplane itself, or the airplane management system, that a bearing on engine number four is experiencing excessive wear and heat, and should be replaced when we land, and then it communicates to another machine, and so it should place an order, and magically an engineer appears on flight at the airport and is prepared to replace it during the fueling.
And the scale that is so big and so large that we’ve had to go into what I refer to it, via access to kind of cope with it. Because a new world, and welcome to the new world, a new world of everything that we use being connected; once upon a time it was satellites and network devices, now it’s mobile devices and our laptops and tablets and phones, and even my brand new Audi has a sign built into it, and it reports constantly on its own health, but also updates itself, and knows where it is, and what maps are applicable, and even tells me when to go a different route if there’s traffic on the road ahead.
Everything that we’re building now, everything we’re talking to you now, is being designed to connect and connect to other things, not just from me to system, but from system to system, and to be able to cope with that we’re having to apply very different thinking at the infrastructure layer, both at the hardware and at the software, and particularly the database layers that systems need to underpin this, and in many ways the database has become the engine, and the apps are really just little bots that do things.
I’m going to wrap up quickly here with this slightly humorous view on kind of where we’re going with these things, and what I refer to as “IoT at the push of a button.” There’s been a new gadget created called the Amazon Dash Button, and this is a little thumb-sized gadget. In fact in many ways, it’s the same as my USB thumb drive. When you buy this thing, it’s about $4.99 U.S. online from Amazon, it gets shipped to you, you configure it with your mobile phone and you literately just attach it to one of your devices, such as a fridge or a washing machine or whatever. In your washing machine example, if you eventually run out of washing powder, you can push that button and it will dial home and automatically order more for you, and magically more will get shipped to you via our good friends at Amazon.
For me, this frightens me, because it’s going to see an explosion of a number of things that are connected on the network and attempting to create connectivity, and generate demand. If you can imagine, one or two of these things is maybe not so scary, but last time I looked, there was over 110 of these things branded, so almost every brand on the planet is going to try and get their own little push-button IoT, that you go home and you push a button and it says, “Order me a pizza.” You push another button and it orders a pre-built lunch for your children for school tomorrow.
That is driving such a massive demand for transformation at the back end, at the application level, in particular at the database level, that I think we’ve only just seen the tip of the iceberg of the type of performance transformation we need to see. And with that, I’m going to hand it to Doctor Robin Bloor and get his insights into kind of where we’re at, as well.
Rebecca Jozwiak: Okay Robin, I have passed you the ball.
Robin Bloor: Isn't that good? Okay, here we go, it’s me. I saw Dez's presentation before I came to this one, so I’d say things that are complimentary rather than just repeat some of the things that Dez said. I thought I’d talk about database evolution in terms of what’s actually happened to the architecture, and so on and so forth, of databases from a historical perspective.
The fundamental problem that any database vendor has is maintaining a flexible architecture that scales and keeps pace with hardware evolution. I’ll talk thought this, but when you actually look back and see the way the databases used to be built, and the way that they’re built now, they’re actually significantly different to what I’d call the architectural design level. It’s worth just reviewing why that is, or at least I think it is. The hardware factors, and Dez has given us a particularly good rundown of the lower layers in terms of memory and disk. What we’ve got now, and this is the future coming, Intel is next, CP who is going to have a FPGA on it. What people are going to do with that, I haven't got a clue. AMD is merging CPUs and GPUs and what a difference is that going to make? These are the kinds of changes that are actually going to make difference to database, and I suspect that Aerospike among others, because Aerospike is driven by performance, it’s probably already taking a look at that and working out where it thinks it’s actually going to go with the way that the product works.
We’ve got a system on a chip that has not taken off yet. SSDs we know about, but the point to make is that they’re actually increasing in speed, roughly Moore’s law’s rate, a factor of 10 every six years. But Intel is about to release 3D cross point, which claims to be able to go more than a hundred times faster than SSDs, in fact, kind of drops into the mix, then that’s going to change the speed at which products like Aerospike can actually go.
Then we’ve got the parallel hardware architectures, in other words the way that we’ve constructed hardware in the sense of – originally it was just a CPU sitting over memory, which sat over disk, but it’s become way more complicated than that. The idea of a system on a chip is that you can actually have parallelism chip to chip to chip and make everything go at extraordinary speed, and we’ve no idea exactly which of these products are actually going to dominate.
That’s just a look at the future, but at the hardware level the performance is accelerating and the costs continue to fall, kind of along the lines that Dez was describing. Your CPUs don't necessarily get cheaper, they just get a [inaudible] faster and so on.
From the business perspective, in some situations, and these are market situations, being first is where the business value is. If you particularly – if you are absolutely convinced a particular stock is going to fall in price, the first person that gets the sell order in gets the best price. It’s really that simple. Therefore, there’s a technology race that goes on to automated trading in the banks to actually try and win these situations. What happened after that? What happens after the banks have done their thing with all of that? You’re suddenly starting to see other areas getting infected with the same kind of needs for speed.
Really what was happening, is the human beings were being removed from the equation, and that happened with internet advertising very quickly. But the thing was, it’s not the specific transaction, the execution of methods, this is a whole business process, it’s the fact that a webpage has just been thrown off, and a decision needs to be made which may be a fairly complicated decision, as to what advert to actually put on that webpage, deducing from whoever the user of the browser is what would be the most appropriate ad to put that on, and so on and so forth. It’s become a very complex thing, and I’ll mention it that again.
But the point is that the performance and scalability of business process, is not the same problem as performance and scalability of a query capability, and this is something that I’m well aware, because of a recent briefing room we did with Aerospike that they’re also aware of. Another thing, when you’re actually working at these speeds, asset properties matter for a transaction, any event processing. They really, really matter. So an awful lot of what some databases are doing, which is losing a letter or two from asset, may work reasonably well in the context – this will work well in the context we’re talking about. It’s not really acceptable, to be honest.
From a technology perspective, you’re actually looking at – I know there’s two kinds of leverage, in order to create the kind of architectures that are actually required to give the kind of speeds that can do, like Aerospike, can do a million transactions per second. You need to actually be very precise in terms of the software development. You can’t just hack away. You need to be concerned about code path lengths. You need to make excellent use in memory, and you’re actually optimizing whole transactions. You need intelligent parallelism and you also need fail-safe parallelism. You need to scale up, rather than scale out, because as soon as you involve the network in anything, it becomes the most likely pointer which you’re going to hit latency, and it’s going to start making the transactions too slow.
You have to get as much as possible onto any given known of a network before you actually scale out, and you really don't want to scale out quickly, you really don't want many processes. You want a network that isn’t being used by anyone else. And you want to have an incredibly fast network.
Accelerated SSD storage is something – actually I think most of this applies to what Aerospike does. One of the interesting things is, is it’s a NoSQL database. It used to be believed – I don't know, a number of years ago – it used to be believed the relational database was the only database and it dominated everything, and it was only this odd little niche situations where you didn’t need to go relational. It’s kind of turned on its head now. It’s the fast databases that are on those SQL databases, and one of the reasons for that, the major reason for that, is they avoid joining data, they store data pretty much in an object fashion. When you’re finished with an object you just store it and then you pull the whole object back, it’s not joining things together in order to actually process them. This is what speed is about. These kinds of techniques that generate speed within the database context.
This is the trail of tears, this is the, what happened to database. The story or the narrative of the relational databases was end of a database actually was not true. Even when they started to come to dominance, it was still necessary. Object databases did the past transactions in those days, because relational databases actually couldn’t do them, and then it turned out that the relational databases using row stores, they couldn’t do fast queries either, you needed column stores. And then we discovered that if you actually wanted to do graphical queries on data, neither a column store nor a relational database would be any good, and you actually needed to have a specifically graph-aware database built for you. Then RDF databases came in, and as soon as you actually started to consider the meaning of semantics and we got the NoSQL databases in, very, very specifically for speed. To call them NoSQL is almost as if you’re branding all of these databases as if they were the same, actually they’re radically different in what lies underneath. The only reason that they bear the name NoSQL is that they don't give a damn about SQL because it’s too expensive. The transaction latencies that they need.
The IoT – which I thought I’d finish on the same point that Dez finished it on – it ain’t over, all of this situation in terms of speed and the latency requirements, it ain't over until the fat lady starts to disgorge this data, and it hasn’t really started yet. A lot of that data is going to want to have the latencies that I’ve been kind of indicating, so I think that’s all I’ve got to say. Let’s hand it on to Aerospike and Brian Bulkowski.
Brian Bulkowski: Hi, thanks a lot for joining the Bloor Group and myself for this presentation today. In thinking about what Dez and Robin were just talking about, I’d like to tell you a bit about the trail Aerospike has taken in providing new database technology and NoSQL databases technology to a number of industries. It’s been a great path. We started Aerospike in 2008 seeing a lot of the trends that Dez and Robin have mentioned. Specifically about in-memory databases being able to take advantage of flash, as well as the kind of scale-out cloud systems, and the kinds of scale required to do personalization, behavior analytics and the kind of celebrity VIP experiences that were discussed.
When we approached the problem of a database that was a front-end operational database that was capable of providing the underpinnings to applications that could be written to solve these, we started with the problem of how could we build essentially a distributed hash table, memory-distributed hash table that was astonishingly fast and capable of things like millions of transactions per second, but at a reasonable price. When we finished our prototype, we realized that then we would have to figure out who might need this kind of speed. Being a Silicon Valley company, we quickly found that it was really the advertising industry that was capable of consuming this kind of information and was interested in it, and so I’d like to spend a second talking about real-time bidding and how this market works.
Robin mentioned how financial trading works, which is the first transaction is often the winning transaction, and there’s essentially a time to market of latency and a value to latency. The advertising industry is slightly different, in an interesting way, because the goal in advertising is a particular – what’s called an impression, the ability to deliver an ad – is an auction and that auction runs in between ten milliseconds to fifty milliseconds. The name of the game, and there’s often hundreds of companies now bidding in real time on every single ad that’s placed on the internet, is to get the most amount of data and bring the best algorithms to bear within that ten to fifty milliseconds over the largest amount of data.
This change and shift was happening in the advertising industry, in every one of those little milliseconds, have a time-bounded complication with the best algorithms over the greatest amount of data, and to do that you’re bringing together lots of small pieces of data. Recent IP address information, recent information about a particular device category, recent information about website behavior, recent search terms, all would go into the secret sauce of a particular company’s algorithms to determine a price and a bid.
This has been a fascinating market to be a part of. We first did our first deployment at Aerospike in 2010 with some of the first companies working seriously within the real-time bidding economy, and then have achieved, basically being that front-end store of behavioral data, for the majority of the companies in that space. What we’ve found since then, and is a particular architecture that I’ll detail through the course of this presentation, is that was all happening in 2010, 2011, 2013 and continues to evolve. Advertising is a very dynamic market.
But that kind of VIP experience, you can think of as placing the right ad, placing not an ad for say child’s products, because I don't happen to have any children, so I’m going to not have an effective ad if it’s placed on that, but if it’s about fast cars that’s the kind of ad to place to Brian. That’s really the kind of VIP experience in deals, in whether to discount or not, if you’re on a retail site, even in fraud detection. Is this the normal pattern of a particular person, or a particular credit card? All of that form of technology of real-time analytics, of behavioral prediction, of predictive analytics, is now seeping out of the advertising industry, which has been doing it for fun and profit now for quite a few years, and really coming into retail and banking, and fraud detection, etc., through a particular architecture. So Aerospike has been privileged to be a part of a number of those cases.
The architecture that we see working, and being practical for doing this, is one where instead of creating a set of queries from an application server, instead moving more of your computation to the app server itself, and then using a database as essentially a storage engine for the kind of objects that Robin was talking about. In this case, these architectures, first of all don't confuse this with your actual analytics here. You see on the right-hand side of this slide that there’s still an analytics here for generating insights. These are jobs that are often working over petabytes, tens of petabytes of data, even exabytes in the cases of some of our large customers, using a variety of technologies. You need to have a big data team, an analytics team, a quantitative team back there figuring out what, say, geospatial coordinates matter, what models work in terms of finding those relationships and creating the VIP experience. That’s a whole problem unto itself and not one that Aerospike has directly participated in, and there’s a bunch of great technology when you’re dealing with that kind of system.
What we’ve been excited about and working with the industry about is, once you have those insights, how do you engage in the kind of machine-to-machine or fast machine-to-human transaction, where you take those insights and make them real for every person, moment by moment? The architecture that we’ve seen using that is one where there’s an application server that’s written and it’s doing all of that math and looking through the models that you’ve created, and looking at recent behavior and doing that over essentially a key paradigm or at least very query-light kind of system.
When you’re dealing with the kinds of data types that we’re talking about, the kind of flows that we’re talking about, with millions of writes per second, millions of reads per second, millions and hundreds and thousands of decisions per second, building complex indexes, multidimensional indexes, simply doesn’t work very well, it isn’t scalable. The way to achieve this form of scale is to engage a lot of parallelism. We’ll talk a little bit about how we do that later. But part of that is a stateless app server written in your own language.
What we often see is a particular project assuming a new application framework based on the people who work there, the technology that they’re using, and the problem that they’re approaching. We’ve seen people using Python, a lot of people use Java, we still see C programmers, because a lot of this is still high performance, maybe even using things like the old MATLAB libraries. And they need to touch thousands upon thousands of data points per second in order to make an effective decision.
One question that I have had asked sometimes is, “Well, Brian, if you’re capable of millions of transactions per second, who needs that?” If you look at, for example, North American payment processing, and Aerospike is involved with solutions doing fraud detection within that system, and supporting application writers who are doing some very innovative things in fraud detection, there’s only a few thousand payment transactions per second flowing through even the largest of payment processors. And yet, when the first company came to us and said they were looking at using NoSQL, and wanted to see what our solution would look like underpinning their application, they said they wanted to touch 5,000 pieces of data in a 750 millisecond window. Well now suddenly you have a few hundred business transactions and a few thousand pieces of data to consider in each computation, and now you’re up in the area of needing millions of transactions per second.
The case of – putting aside advertising for a second, the case of fraud is fascinating because where there’s money, there’s fraud, and real-time prevention of fraud, as opposed to trying to sort through analytically after a fraud has happened, is really a matter of bringing online as much data as possible, and you can think of it as a reflection of that VIP experience. Is this person behaving in a way that they do not usually behave? And thus, the chances of it being a fraudulent system, and not actually this person, goes up. Does this person usually access through a particular device or set of devices, with a certain set of screen resolutions? Do they usually exhibit a particular behavioral shopping pattern? Perhaps we can nip fraud in the bud during the course of the transaction itself. That should remind you very much of the kind of thing that happens within a transaction in the advertising system.
The kinds of systems we solve are ones where each individual payment processor has a big data team, they have a lot of historical data, they’re creating new models, they don't share with us at Aerospike all of the models, because they’re really a secret sauce. If you’re a subscriber to Gartner and you heard Gartner talk about the algorithm economy, this is one algorithm and one company fighting head to head in order to drive down fraud and to bring up the number of successful transactions, because you also don't want to block transactions. That’s the kind of projects that we look for in Aerospike at these levels of scale.
Another case that we’ve been working on with financial services companies, is what’s called the Intraday System of Record. In this case, what’s happening is, the kind of richer experience, even in a retail trading system, is one where I want to be able to look at my particular position and I want to do so extremely accurately. I don't want to have a catch in front of my DB2 system. Instead, I want to look at the exact data, and between mobile, but also things like a risk recalculations, risk recalculations should now be done on a minute-by-minute basis, you want to be able to recalculate everyone's risk as well as the global risk, systemic risk across the entire company within a few minutes.
And again, it’s the same problem. Every single account that’s a particular, think of it as a key value lookup to a particular object, then this can be done in parallel, and most importantly, this paradigm allows you to write your code and your algorithms in a high-level language, which is easier to debug and faster time to market. In this algorithm economy, I need to be able to get my algorithms online now. This is a very different problem for modeling and business relationship, which is what relational systems are great at. When you have a table of parts, and those parts are associated with orders, and those orders are associated with people, you’ve got a business process that can be strictly modeled and probably won't change for the lifetime of your business. However, a new algorithm to find new fraud pattern has to be written accurately and quickly, and gotten online, making business decisions within a matter of days at the very least, if not faster. A NoSQL solution for this kind of system of record is really an amazing system for these guys, because it allows them to ingest data very quickly, as well as to build new algorithms, so not just a new customer experience in addressing mobile, but really building out a wide variety of new applications.
What we see in the long term at Aerospike is the fact that each database type, each physical layout of data on disk has its own components, and at Aerospike we’re really focused on this key value or role-oriented system, as Robin said, with high transactional consistency, and really allow people like column stores and high-volume data lakes and as well as hardcore transactional systems that have had reporting constraints on them as well. We see all of them needing to feed into a variety of different query engines. We see some of the JSON-based query engines. We see things like elastic search, we see Spark, all needing different varieties at different times of things like column stores, as well as row stores, which is where Aerospike excels.
We really see that these different types and the industry is getting to a point where picking the best of breed of each one of these is going to be a necessity. Unfortunately, due to the reality of long-term analytics and batched jobs verses analytics, and operational constraints, we probably won’t get to the point of having a single, one size fits all, but we will get to the point of being able to choose clearly between some of the core data layouts.
Let's talk for a minute about the innovation of flash. I do still get the question, even though as was commented earlier, flash has been with us now for a long time. When we started Aerospike in 2009 was when, I believe 2009, maybe, yeah, 2009 was when Intel came out with the X25, which was really the first mass-market SATA manned flash drive, and there were a number of flash systems before that, but really that was the one that broke into a lot of technology’s consciousness. Fusion-io really brought flash to the broader enterprise market after that.
What’s happening now is the advent of a system called NVMe. NVMe is a standard similar to SATA or SAS or even SCSI that allows different card vendors to interoperate with drivers within the operating system at a high level of efficiency. So it’s creating a greater level of performance, first of all because NVMe is based on PCIE as its underlying transport, which is much faster than SATA, SAS or anything else, but also it allows best-of-breed drivers.
For example within Linux there’s this guy Jens, and Jens is the NVMe driver guide, Jens expo, and he’s doing a better job than any individual tn Intel or Fusion-io could have done with their individual driver, with all their resources. When you have the power of the operating system itself being able to build the best driver, we’re seeing some really amazing levels of performance. This all backs up the idea that flash really can provide a lot of the low latency of RAM.
Now, Aerospike is still a great RAM database due to its cluster model, however, we find that once you’re doing a network hop, which you need to have scalable storage, you’re already spending at least five to 50 microseconds, the extra 70 microseconds of NAND is usually not an impediment, and you might as well use flash, given that NAND flash, given that the network is already involved in that. Many people then wonder about how – this all sounds great if you’re buying your own hardware, how are the public clouds doing? I think you’ll find right now, no matter what public cloud you’re using, those public clouds have very strong flash offerings. It does differ a bit from cloud provider to cloud provider. Amazon has its I2 instances that have been out for I think a year, two years now, that are really pretty high quality flash devises, and Aerospike has the deployment pattern on top of them.
I’d like to call out Google Compute, Google Compute Engine, Google Cloud specifically, because in our experience they so far have some of the highest performance devices and some of the most flexibility in terms of deployment patterns. But also you see new deployment patterns like Pivotal, which is a sort of public/private, so you can do right Pivotal apps both places that support flash and support different storage devices as well as Docker patterns. So really, this is a point in history where flash is not only available for you to buy and put into your data centers, but really has sunk throughout all of the infrastructure providers, because it’s really the best way to get high-IOPS systems at a very reasonable latency.
Just one moment about Aerospike – Aerospike is a cluster distributed database, which makes it very amenable for cloud-style deployments as well as data centers. We find that the flexibility of being able to add more data and more performance is absolutely necessary in these kinds of net new applications because you start a project, you don't know if you need fifty thousand transactions per second, a hundred thousand, a million, two million, so you want to give yourself some headroom of being able to add servers. And yet, you want to scale up so that each server is able to be fast on its own. You don't really want to end up with five hundred or a thousand servers that are database servers that are slow. Scale out is not the only game in town, its scale out and scale up, as Dez was saying earlier, there’s a new Z axis.
Hopefully that gives you some new ideas about how speed and scale is addressing new markets and perhaps there are projects that you’re working on where you’ll be able to consider really building out more rich applications and using an application framework with a more key value or NoSQL database below it. At Aerospike I’ve certainly seen a lot of our customers and a lot of our open source users succeed with that pattern, and I look forward to the industry adopting it to a greater extent.
Rebecca Jozwiak: Thanks so much Brian, and I’m sure Dez and Robin have some good questions for you. Robin?
Dez Blanchfield: I’m happy to jump in. Robin, do you have a question? Otherwise I have a quick one I can start.
Robin Bloor: Sorry, I was on mute. I did dive in, but nobody heard me. The question immediately occurred to me, because this is a very sophisticated set of technology capabilities. In terms of the existing customers you’ve got, what’s the kind of escalation or transaction rate that you’re experiencing concerning some of these ad applications? Is the transaction rate continuing to rise? And if so, at what kind of rate?
Brian Bulkowski: Interesting question, Robin. Each industry has its own curve in each company. Let’s take North American advertising, in say 2012, North American advertising was running probably closer to 200,000 ads per second, in sort of standard intraday, not my time, and it’s now escalated probably to about three to five million ads per second. But then an interesting thing happened. The ad industry started addressing some fraud concerns, and the parts of the industry that are able to block fraud, saw transaction rates drop a bit, about a factor of two, within some of our more sophisticated customers that were able to determine fraud. Of course they had to do some database lookups in order to block fraud, so it sort of ends up being sort of the same in the end.
An interesting use case is within telecom, I didn’t really mention that, telecom were seeing transactions increase due to billing based on every single packet that passes across the cell phone network. In the old days, we had call detailed records and once a minute, a call, what you know, a little ping would go through the network and does this guy still have a minute left? Now we have to build and even route based on every packet on the internet. That’s a – sorry within a mobile network, which is suddenly now millions of packets per second and something that’s growing over and over again. So one case is every application is driving a nice little sort of 2X per year. Within some customers, we see, “But wait, I have a new application. I want to add some fraud to my risk. I want to add some deeper customer experience to my fraud and my risk.” Each one of them creates new load on the underlying database.
Robin Bloor: Yeah, I mean I think that was what I was hinting at in the brief presentation that I gave, that these – we used to think a transaction is, somebody does something and maybe there’s a cascade of events and it all gets recorded, and now a lot of transactions have an immense amount of lookup, and you gave some examples in the presentation. And therefore you’re not actually executing a transaction anymore, you’re actually executing a kind of application that can have many, many elements to it.
The other question before I hand over to Dez – because we’re obviously tag teaming on this – the other question that I would like you to answer if you’ve got a reasonable answer to it, is both Dez and I expect the Internet of Things, or the Internet of Everything as it’s sometimes called, to create a fairly dramatic amount of transactional traffic. Can you speak to that? Is that your experience, have you got customers coming to you with that particular kind of problem, and what’s your view on this at the moment?
Brian Bulkowski: Sure, I think there’s a little confusion, and that’s to put it mildly, about Internet of Things. The customers that I see so far are simply bringing the internet to the things that they have. Think about those Amazon buttons – it’s all Amazon – those buttons, you can’t repurpose them and have them go to Walmart online. It’s not like a browser that you can mix and match everything. On the other hand, machine-to-machine is happening, and when you plug in your Tesla car to charge it, Tesla sends a huge backflow of information, every single sensor into the car, but it flows into Tesla’s computer for analysis and improved quality. What I see is, all of that machine-to-machine, and all of the sensors within an individual company, creating new demands.
Now mostly today, that’s flowing into these analytic systems, and take the case of Tesla; Tesla’s first use of that, to my understanding, was to improve battery life, under “What operational temperatures are they, what are the loads? Let’s look at it, let’s design a better battery.” But then they start thinking, and that’s all great, that’s sort of a deep analytics problem that’s fascinating, the next question is, “How do I improve the moment-by-moment experience?”
Now let’s take the case like Nest, where you’re trying to do predictive analytics to change a home’s temperature moment by moment. That’s the kind of case where we start seeing in Aerospike, where there’s this huge data lake and there’s this huge analytic processes, but what am I going to do now? I’m going to need to keep, think of it like the cash, some portion of the last week, the last month, maybe even just the last day's worth of information, probably on a back end because we’re dealing with simple sensor devices, and I’m going to be doing a set of analytics on that moment by moment to change experiences. That kind of Nest-like experiences, one that I see Aerospike use cases for.
Robin Bloor: Okay, the thing that I was expecting with the Internet of Things, was that you would start to get threshold triggers and that they would start creating cascades of events. Have you seen anything like that, or is that not anything you’ve seen yet?
Brian Bulkowski: Dez and I were – I was just asking Dez's opinion on that when we were pre-show chatting. What I have not yet seen is the kind of cascade of one company’s data cascading into another company, that my Samsung fridge is talking to my LG washing machine because it just figured out that I spilled a whole bunch of chocolate all over the floor, so that kind of company to company device by device, I think I’m still waiting for that in terms of Internet of Things. I think there’s some problems in business and security that are mostly non-technical that need to be answered in order to see that.
Robin Bloor: Okay, Dez?
Dez Blanchfield: I have some very strong views on that particular last point actually, that I just briefly will bring into the conversation. I think often business and technology think that they actually drive where the demand is coming from, but when we look at what happened when the iPhone became a thing, and in my mind it was sort of the first mobile device, if you’ll pardon the pun, but a device that could be carried around that can actually run lots of little apps in your pocket, and it brought about a significant transformation on what we thought about being a computer. A lot of people think about iPhones or smartphones, or Android phones as phones, but they’re not, they’re actually just a little computer that runs apps, and one of the apps it runs makes calls, and they’re not the calls that we think of anymore, they’re not an analog point-to-point call as Brian highlighted, they’re little packets that get routed around.
But more often than not, what we’ve seen is this insurgence of smartphones actually not really being used to make calls that often, probability 98% of what I do on my smartphone is not make calls. It’s everything but calls, it’s apps. I think this cascading effect – and I’m keen to bring this to a question quickly – but the cascading effect is actually brought about by consumers, and in fact I have this one liner that I throw out quite often to get a bunch of CXOs sitting up in the room and paying attention if I think they’re falling asleep with the presentation I’m doing, which doesn’t happen too often, hopefully.
I sort of said it in that disruption that you’re seeing in your business is actually not being driven by technology exclusively, it’s more often than not being driven by your customers. And they sort of sit up and actually wonder, what does he mean there? So when I think about the use of technology, I mean we saw USENET, we saw all these kinds of fun things happening on the internet, but not many people predicted social, and the impact of it. Everybody wanting to tell everybody what they had for breakfast, and the noise that that created and the backend technology we had, and then of course advertising is trying to fill it up with things.
I think we are going to see a cascading effect to a point where devices are talking to devices, consumers are just catching up with what that actually means, and what that can do. You raised an interesting point around why Amazon button won’t talk to Walmart. I’m going to post this question, what happens when Walmart gets their own button, and then what about if the top twenty Amazons and Walmarts and other major distribution and retail networks all get their own buttons? Where does that take us? Specifically, my question with Brian is going to be, “Where are we going with this whole new paradigm of performance? You’re at the bleeding edge of it, and you’re working with companies that are doing it at both the physical infrastructure level as well as the transferring data level. Where is this taking us, when this next big wave comes? What sort of insight can you share around that with what’s happening at the backend from your experience?”
Brian Bulkowski: Sure, the way I think about a lot of these things are to focus on the user experiences and exactly what you said, it’s the users that drive, even though, as technologists and as business people, we might come up with a clever idea that we think the users like, and I’ll sort of go back to the Nest example. When my sister installed Nest in her house, she said, “My house is quieter, I can hear things. It’s not even just that I’m paying less for power,” she is, but you could now not rip that Nest out of her hands because she likes being in a quieter house as opposed to one where the heating is blowing on at a maximum and then turning back off.
The question ends up being, what are the user experiences that we can empower? That ends up being, that quality-of-life experience, that if we have the money and we’re in the first world, we would pay a lot for. I’ll give you an example from my own house, my girlfriend likes cold milk. She likes really cold milk, and so often we have to try and figure out where in the fridge is going to be cold enough, and not have the rest of the things overheat. Well this is a great – and I said to my girlfriend, “Would you pay $10 a month to have cold milk and not to have frozen cold cuts?” She was like, “Absolutely.” And getting $10 a month out of any consumer is tough.
I think that in these experiences we really have to keep an eye on what is that consumer-end experience that really could be driven. I think that was part of the secret of the iPhone. I think it’s part of the secret of Tesla building a better car with all of the data, abolishing the idea of a product cycle and a yearly release and doing continuous improvements on every part. We’re going to have to come up with some clever ideas on how to actually use all of this data in a way that’s compelling moment by moment to people’s lives.
Dez Blanchfield: Yeah, that’s great insight. Leading on from that, the other end of the spectrum, that echoes exactly with the sorts of things we’re seeing now with what consumers are asking for, and all of us have something in the house who cold of this and warm of that. The other end of the spectrum is then, and we’ve seen this in sort of the traditional “big data world” where data assignments are becoming rarer than hen’s teeth and those that are on the market are being offered more than the CIOs are earning in some cases, the types of companies you’re working with and the types of development you have seen, is it the case that the types of developer and the type of data architect and the networking specials, are they becoming harder and harder to find? Do we need organizations to start thinking now about getting ahead of the curve of the type of skill set they need in the back end for the type of developers, and data architects? What are you seeing at that level as far as the skill resources that they will understand how to put this technology into good use now looking like?
Brian Bulkowski: Yeah, I think that is one of the challenges facing the organizations I’ve talked to. Whether it be a – the worst problems I’ve heard about are actually sort of larger enterprises, because if you say, “I’m from this large bank, I’m from Chase and I was a data architect,” then you’ve got the world’s your oyster and your salary goes way up, so there’s this churn problem of getting a job in one of those places because there’s not enough people, and then being able to just move from job to job. I hear nothing but that kind of problem, and that’s actually one of the reasons why I’ve been focusing Aerospike around using tooling that is appropriate for the particular project team.
Instead of trying to walk into a project team and say, “Hey, you should use our query language.” Look, if those guys, they’re driving the bus these days, guys and gals, and if they use a particular query language and tooling, they’re going to stick with that, and I can’t talk them into anything else. My goal is to be able to put the kind of Aerospike power as a database behind whatever tooling they’re using and that’s part of this idea, the slides you’re seeing about the Poliglot database future. I need to support the patterns of application and analytics between these guys, because it really is difficult trying to find people who have the mathematical background as well as the statistical capabilities to navigate this world.
Dez Blanchfield: Another interesting thing that people may not be aware of, I mean Aerospike is a very strong player in the open-source world, I’m keen to get a very quick insight to kind of what that means as far as how the business operates and what it does for you. You mentioned you worked directly with folk who are doing things right down to the kernel level inside, so the Linux kernel. There are some big players that are in this space, and there’s some famous brands that we won’t mention, but an organization like Aerospike, in your more modern recent history, the open-source experience, how does that fit into the big picture and what competitive advantages have you seen that give you?
Brian Bulkowski: Sure, when we transitioned to open source in 2014, we did it because we realized that a core infrastructure, like a database needs to be source available, it needs to be trusted and a natural counter balance between the old world of closed source, and once you invest in a particular database, those guys have you at their mercy for technology cycle after technology cycle, and there has to be a balance. We need to be able to bring out versions that do new things, and maybe that’s in an enterprise version, we need to have a dual-license model that has an open-source version for people who are kicking the tires who are doing nonprofit work, as well as an enterprise version that is a proprietor license and allows unlimited work.
And of course we’ll also have the highest levels of speed and scale, being an enterprise version. We believe in the duel-license model, and that’s been great for our business. We want people to get started with Aerospike, we want small projects to kick the tires, it’s super easy to just go to Amazon, launch a confirmation script and have an Aerospike cluster running within five minutes. On the other hand, we want to give more to the enterprise customers.
Dez Blanchfield: We’re kind of getting close to the top of the hour, so I’m going to pass back to Rebecca in a moment, but if there was just a one liner that you would throw out there, sort of the advice you would give to folk who are looking to get into the space of the technology you’ve brought to the market and how they’re going to adopt it, what would you say the first step for them is to sort of at least dip their toe and start looking at how they’re going to get a competitive advantage from your platform?
Brian Bulkowski: Sure, part of the message here is that there’s levels of speed and skill that are now easy. You don't need a thousand-node Cassandra cluster to achieve millions of transactions per second. You can do it even in the first phases of your project. So things are a lot easier than they used to be. Then the second piece of advice is you are going to have to come up with, just as you’re saying, math business process customer engagement models that make use of all of this data, so the good news is the data is available, the bad news is you actually have to go find some patterns and some compelling use cases.
Dez Blanchfield: Yeah, great advice, so I’m going to hand back to Rebecca now. Thank you so much for that, it was a great little chat about the technology, I appreciate it.
Rebecca Jozwiak: Thanks, Dez. I do have a couple of good questions from the audience. Let me throw up this slide. I know you talked about the system of record and mainframe stuff, but how often are you seeing absolute offloading or is the replication an end-of-a-day reconciliation, kind of what you see more of?
Brian Bulkowski: What we see in Aerospike is using a NoSQL database in front of that end-of-day reconciliation system. You need intraday, the correct answer. You can’t have the wrong answer, and that was what Robin said about asset is underappreciated, but the business processes around the legal requirements of reconciliation can get quite complicated and there’s decades of technology and decades of law and law practice around doing reconciliation. So what we see at Aerospike is, you’re going to be doing your algorithms on a hotter database with more transactions per second. But for legal reasons, you absolutely need a reconciliation system that has been through those legal processes. We see both, and we see that this is essentially the two-tier IT practice as exposed by people like Anderson Consulting and Gartner to some extent. We see a lot of that.
Rebecca Jozwiak: Okay, good. Someone else showed interest in this particular slide, he said that it was really interesting and wondered if you could just go into a little more comparing flash versus in-memory.
Brian Bulkowski: Sure, well let me take a quick side bar, again, I know we’re close to the end of time. Well flash is memory – it’s chips – I tend to think about RAM. So RAM has particular characteristics, requires a lot of power, it’s very good at random writes as well as random reads. Where NAND is capable of fast random reads and lower power, but it’s very bad at random writes. There’s some subtle differences in how these two chips operate at the lithography level, that create a number of technical differences.
In the case where you’re doing analytics and you have to skip over a lot of data, or in the Aerospike's case, where you got indexes, indexes are still very good to use in RAM because of parallelism and random access. A higher level of random access is required. In Aerospike though, we find using those indexes to find a particular object or chunk of data, that’s the appropriate place to reach out to a NAND because it becomes sort of a larger store underneath the indexes. That then is one transaction to a storage device, but still after doing a lot of potentialities and filters within your indexing system.
Rebecca Jozwiak: Okay, good. And then, I know we talked a lot about the IoT already and one attendee comment said IoT is largely beneficial, but are companies, government entities and developers growing securely and securing data at the same rate, do you think?
Brian Bulkowski: Maybe Dez, would you like to jump in?
Dez Blanchfield: Yes, I’m happy to jump into that one. I think the answer is no. In fact, one of my favorite throwaway lines on this topic very, very briefly is that I think the explosion of machine to machine and general Internet of Things, communication and the security, the risk around it, we’re at the point now where governments can’t keep up with the rate of change. And in fact we know a lot of organizations can’t keep up with the rate of change. In fact, if I paraphrased it, the rate of change today is so great that the organizations are having to sprint just to keep up, but they’re having to sprint in multiple races. I don't think that the law, and I don't think the government in general, either state or federal level, are able to keep up with the rate of change.
Now, my general advice to people is kind of act now and ask for forgiveness later. There’s been many examples of that in the past. They will catch up, but I think it really is now up to business and technology providers to kind of innovate in this space and to ensure that we’re familiar with the security risks or privacy risks and we need to deal with those. Banks in particular, as you mentioned, when you think about what a bank organization has traditionally done with things like anti-money laundering and know your client, the AML/KYC challenge, it used to be that every three to five years we would try and meet compliance.
Now I think that needs to be built into every single transaction. You have always been able to do that at bid level with advertising and stock and bond and equity trade, I think we’re at the point where the performance you’re bringing about with Aerospike platform allows us to now think about how do we bring privacy, how do we bring security into that immediate real-time decision chain? And so the answer is no, I don't think governments are keeping up. I think companies need to keep up, and I think we need to act now and ask for forgiveness later.
Brian Bulkowski: Let me add a couple points as well. The guys I deal with, the technology companies I deal with, are very cognizant of making sure they’re on the right side of the law, and a fair amount of the discussion is, is this PII, can I use this, how am I using this particular chunk of data? What was its providence, and is this a protected decision or experience? How do I do all of that? So that’s the good news. I do wonder sometimes about our discussion as a society around where we’re heading, and if even our society discussion is at the appropriate level in terms of using the new capabilities from IoT all the way up to machine learning, which is the only way to sort through the volumes of data we have. But the good news is, the guys I talked to are really on the right side of trying to do right by the legal decisions we’ve made.
Rebecca Jozwiak: Those are some really good answers from both of you, and I totally agree. I don't think that security is moving at as quicker pace as technology development, particularly when it comes to the Internet of Things, but I have to think that people are doing their best and hopefully we’ll get there. It’s always a little hard to stay ten steps ahead of cyber thieves and cyber criminals, but we’ll get there.
Well folks, we’ve gone eight minutes past the top of the hour. I’d like to thank our guests Brian Bulkowski from Aerospike and Dez Blanchfield and Robin Bloor. Thank you so much. You can always find our archives at insideanalysis.com, SlideShare, YouTube, we’ve got a lot of good webcasts coming up folks, it’s been a busy month. It’s going to be a busy month next month, so stay tuned and we hope to see you next time. Thanks folks, bye bye.