Build Top Notch Marketing Technology at Scale (Cloud Next '18)
Well. Thank you very much for turning up at 9 o'clock I'm. Damian Laura I'm a managing director at, Google cloud, and we were to talk to you about how to build top notch marketing, technology, at scale and a. Little bit about myself I. Almost. 11 years in Google, I spent two years running. Our. Publisher. Network in Europe, then, six years running our double-click ad technology in Europe and I've, been two and a half years in cloud and, I'd. Love to get an idea but who is in the room how. Many of you are actually. Building. Marketing, technology, that you sell, as a SAS solution. Ok. So that's about half, of the room how many of you build marketing, technology, within an enterprise for, a marketing, department our to support marketing department, ok, um how. Many of you are our users, of marketing, technology, yourselves. Anyone. In the wrong session oh. That's. The first time Sony's actually answered that question good man. So. And. This. Is a really really exciting space. The. Number of our. Tech and Mar tech companies, we, talked, to either chief Mar tech or to, luma you know the famous loom escape that is getting like as more, and more companies get launched the size, of everybody's logo gets smaller and smaller so. But um chief. Marketing actually very helpfully provided Google, Doc, with all of the the names of the companies that they actually list in they're impossible to read poster. And it's. Over, 10,000. Marketing, and advertising technology, companies that they've now identified, in the space which. Is is it's pretty amazing how much growth we're, seeing in this area and. It's. It's no surprise because you, know we're still trying to solve those fundamental. Opportunities. And challenges that all marketers, have like how can we understand, our. Customers, better and really personalize their, experiences, how can we measure better across all the channels and drive really really excellent ROI and how, can we unlock the creativity, of our teams so, instead of them having day to day. Work that doesn't Drive a lot of value like freedom ought to really be able to do at, top value work and then of course you've got so many different changes in the, environment you've got like the huge shifts we saw over the last years in consumer behavior towards. Mobile. And. Of course now everybody wants to see how they can use machine learning to really drive efficiency, effectiveness and value, so, am. And. This. Is only getting more important, I think for the enterprise this, is a pretty, amazing stat, that 22% of a company's marketing budget is actually spent on technology. And from. All of the analysts I talked to I think this number, will only go, up because. Of the promise that marketing. Technology, holds for. Everybody and, they. Really want to use there's. A really fascinating, set. Of trends, happening about how, CMOS, and CTOs, and CIOs how that. Is. Converging, there's another stat about how like, CMOS, are actually controlling, more and more of an enterprises technology, budget because. The, enterprise sees how this technology, can actually drive top-line, revenue and, growth, for, the company so, that's. Very very interesting to us here, in Google Cloud obviously not only are we talking to CEOs.
And CIOs but we're now actually looking to talk to the business line folks. And show how Google cloud can drive real. Business impact. For, them and. So. Here's some of the, opportunities. And challenges that we hear, from people like you and, that. We need to address when we're thinking about building marketing. Technology, so obviously. Creating, tools to allow marketers. Your. Customers. To, meet their needs and we. Hear a lot about people wanting to leverage machine learning and. How. To make sense of it on how to take. Action on us and. Unlocking. Innovation, throughout your organization I think that's a really important, part to think of is how do you make things easier, and faster for the people that are actually using the tools so, that they can focus. On what, every, CEO wants, which is focusing, on driving, growth and driving revenue so. As. Competition, increases of both for people who people, in the room who are building marketing, technology tools that you sell, us as a solution or people who are building. Tools and, within a business you're, all focused on wanting to drive innovation, differentiation. For your customers be the internal or external and, the. Sophistication. Of those solutions. Is only growing. So. How, can we help, so. First. Of all you. Know in my journey to cloud one of the things that attracted me to Claire after eight years working on the on the MARTA. Cannot excite at the business in Google was, I had. A hypothesis. That, think. About Google Claire it is it's the externalization, of the technology, that, we use, to. Drive. The largest search and digital marketing company on, the planet so, I wondered. What. That technology. Be really. Good at solving marketing. Technology, and marketing. Analytics problems. Considering. It was actually built for, a marketing technology company and I. Think. That the answer. Is best answered by people, like yourselves but in in, working with well. Over a hundred companies over the last two years I think we've been able to see really really clear examples, of where Google said leap does excel, in this particular space, because. Of the way that our cloud was built but also because of our deep domain knowledge within marketing and I, look at we, have over. Seven and, products. That have over, a billion monthly active users on, our platform, and this has really helped us to understand. The type of innovation. Scaling. Security. Issues, that you're facing every, day and, we want to help that infrastructure, that we've built really.
Support You in your goals and what you're looking to build. So. Let's, talk about three different areas, where we think we. Can really help that, we stand out in so developing, features, and products faster helping. You to scale globally and, efficient. Pricing. Let's. Start off with feature and product development first so. That's. Talking about main ad I'm actually going to talk through some customer, examples here rather than just talking about like, conceptually, how can Google help you so um this. Is a real-time main out is a real-time bidding a programmatic ad retargeting, company. And they're. Originally, running. At 50,000. Requests. Per second but they recognized that they needed to grow faster, um and. In, moving, to Google cloud, the. Important part was actually the. Impact on their business outcome, they were actually able to improve their bidding accuracy, by 75%. And. Obviously being, a, performance. Management company, performance. Marketing company that made a huge difference to both are they were able to provide to their customers and also their own business outcomes. They. Launched a new platform on Google cloud and focused. On AI, machine, learning and and. In doing that they were able to apply real-time. Predictive analytics, to ad retargeting and. Again. A very very significant, impact on their. Conversions. Which, is, big. Part of their business big, part of their growth and how, to make. Money for themselves and for their customers, in. Terms of scaling globally, and I've. On. A number, of occasions actually, shared the stage with Don from infectious media and I think his story is is, pretty fantastic. He actually um he. Pretty gleefully, talks about the fact that he originally built, the. Technology. Infrastructure, on a different cloud and and. Then just, could. Not scale. And. Actually made the decision that III haven't come across like that I just asked the audience, he. Actually moved from cloud, back to on pram, in order to be able to deal, with some of the performance, issues he was saying and I'm just asking it has anyone else done. That have you moved from cloud back to on Prem it's not a direction we see a lot of people. Go um and. OK, we've it. He. Could he explained the logic of it in terms if he did actually solve his performance, problem but all that happened was he then had a gigantic scaling, problem the scaling problem was that his course started scaling linearly back, when he was on pram so, he. Actually hit a point where they really wanted to launch in, a pack in Asia and they, knew that was going to be a huge, huge growth in their business that's the record in this dilemma of whether they were going to launch. On pram, and then migrate to Google cloud or whether, they would actually. Move. To Google cloud and then launch but they knew that there was a huge amount of revenue and business that they if they either made. A mess of it or if they delayed too long so we were able to convince them that they could actually do both in a very very short period of time and. They. It. Went very very well for them they actually migrated her their entire business in a week with, no downtime. Then. Launched in a pack which quadrupled, their business and. One. Of the most amazing stats about that is that at, 4x, the. Size. So, they went from 35,000. Transactions. Per second to four times -, after. Launch in a pack their cost base remain the same compared, to what they'd been on pram so they were absolutely delighted with both the performance, the scalability, and the. Cost, so. Actually. That's the next area, they want to talk about efficient pricing so. It's. Moving. To the cloud is not all about saving. Money and but, we do recognize that efficient pricing is a very very important, and part. Of your, business model so. And. We want to think of it in terms of by. Providing efficient pricing that you only pay for what you use that, gives you more resources that you can reinvest in your business and reinvest, in growth, and. So. Nor, X also. Programmatic. Advertising platform. Also totally. Focused, on scalability, and speed, and. They wanted to build out their real-time bidding functionality. Powered, by machine learning um. Here. They did they moved, across and they got an immediate 30 percent reduction in cost which, they were then able to reinvest. In the growth of their business they. Actually, um. Said. That although that was great that one-time cost reduction was fantastic, what, they appreciated. Even more was their freeing, up their. Precious.
Developer, Resources from, doing operational, tasks, and getting them to actually be able to develop, the, tools and the technologies, that were actually going to drive growth and, revenue for their business so, it's the combination I think of the freeing of resources, that, losing the opportunity cost and then of course there is a cost benefit obviously of moving to the cloud and. So. What does this mean for your business and so. We've built our cloud, to be hyper, efficient. Very. Security-conscious, scalable. For, our selves and we now want to bring that to, you, and help you transform, the way you can build your, own marketing technology. So. This stage I'd like to invite my good friend Sudhir. Husband, who's the PM. Director for data analytics to the stage to go, even deeper on what, we can do, to help you to build out top-notch. Marketing technology thank. You. So. As Damian said I'm a director, of products for data analytics, at. Google I've been here around nine months prior. To this I was VP of engineering, at Zulily I, owned, I built all the data platforms, there I won't. Add tech. Merchandising. Tech product pricing, all of that stuff and I built all of that over three. And a half years at absolutely, so I'm, happy to share some of the experiences, from my past as, part of this session but, before that let's actually go and talk about Google. Cloud and what, are the key areas. That we try to go ahead and help our customers, right one, is the. Whole marketing. Tech is all, about big data your. Massive amounts of data that you can collect leveraging. That data to do better segmentation. To target, right. Customers, reducing, your. You. Know total. Cost of acquiring. Customers getting, better return on investment so it's, all around data analytics, it's all about machine learning and I fundamentally, believe that we have the best technology for, for, those things in in. The world the, other thing is being open, and having. An open. Open.
Source Based technology. Or being an open cloud is super critical because. All customers, do want the. Solutions, that can run in multiple, environments right most. Of our customers are, in a. Hybrid. Environment especially with on-premises, in, cloud on. GCP some of them are across multiple different clouds so it's important to have technologies, that can actually span between different, different. Environments, and then of course one of our key differentiators we, don't talk much about it but one, of our key differentiators I fundamentally believe is our global network the. Reach that we have with our network is pretty, amazing in every part of the world and we'll see a bit of that but. Before that let's talk about the, data and, we, are a data company fundamentally. The whole Google as. A company we collect massive, amounts of data we process that data we, analyze, it and we build some amazing products we damien. Was saying we have seven products that are billion. Plus active users we are actually going to be roughly I think eighth one is coming I heard earlier in, the keynotes with drive enterprise may be reaching. Reaching. Billion user so so, that's what we do we understand. Users. We understand, what they need and we build better products for them and so. So big data is in our DNA one. Of the other key learnings that I have is and. This this is one, of the things I think folks or somebody had the article. On this is if if, your company's not good at analytics, you're never going to be good at AI so. Fundamentally, the first thing I always say is get. Better at core analytics, get better at having good data and then, then. You'll be able to do machine learning pretty easily on top of that and one, of the things that we are doing is trying to get the, machine learning tools make it easier and easier for our audience, to use them and leverage. Them for solving problems and I will touch upon couple of them now. This week that we have launched but, from a core platform, perspective what we're trying to do is four, things one we, want you to focus on analytics, not on infrastructure, so, if you look at some of our products that we are building, they're all server less so you can literally bring your data in you can start analyzing it start, doing segmentation, start doing targeting. Just just, based on the data that you have so you don't need skills. To you don't have to worry about monitoring, products you don't have to worry about managing scaling. Infrastructures that's our problem we will take care of that the. Second thing is comprehensive, solutions I will touch upon the whole portfolio of products we have and how, they fit into what. You can do with marketing tech I will touch upon that more details, now, we have an end-to-end machine learning lifecycle, so if you think about it. 80%. Of the time our data scientists, and analysts spent on is cleaning data and 20%. Goes into actually doing real. Analytics, or real machine learning stuff so we're, trying to build tools that are making it easy for you to go do that now that whole lifecycle and then of course innovation, we are innovating faster we are delivering newer and newer products, so, this is a whole portfolio of products I won't go into everything here the, key point I wanted to and this one is we. Like if you wanted to collect real-time, events from globally, wherever, your customers, are and then, you want to take those events as soon, as a person comes on your website on your mobile device and you want to do a targeted. Offer. For them or something we, have products across the board so if you see cloud, pub/sub it's, a globally, scalable event. Collection system where any kind of a clickstream event or something you can just collect and then, take action on data. Flow data proc their, large scale of the, data processing. Solutions. Where you can get that those events in real time figure or I will give a real example and Zulily we did this where, we collected every time, poor. People came into the website we collected those events in clickstream collected. Them in a messaging system real-time. Aggregated that information, and she just showed it back you see this on different, hotel sites and all n number of people are watching this room. Right now or Y number of people are buying this stuff now we just did that it took us like couple of weeks because we are all pipelines were already in place immediate. Uplift in our conversion, rates because, that's, an instant feedback for customers, what they are doing so those, kind of systems are really easy to build with now, with things like pub/sub data flow rate up rock and, then bigquery is our large scale data warehouse we'll talk more about it where, it allows you to go ahead and build.
Large. Collect, large amounts of data and actually, process it and, do. Segmentation and, things like that and then if you want to get into advanced analytics, you can use ml engine tensorflow, for like real, machine. Learning at scale, here. Is our database portfolio, again I'm not going to do a lot of talk, on the on. Specific, products but the most important, thing I would say is as, I said bigquery large-scale, data analytics, you can do in that the, other side of thing is if you're looking for a globally. Scalable, relational. Database which can be consistent. Across the globe for for your transactions, spanner, is a solution like spanner can. Scale globally we lot of our internal systems actually leverage spanner for the consistency, across the globe so, that's one of the products that you should absolutely look at but, let's talk about bigquery, for a second because that's, where we have key. Innovations so lot. Of internal. Analytics that, we run in Google, all runs or on something called dremel that's a dremel paper externally, bigquery. Is an externalization, of the same technology, so, first of all you will use the same petabyte, scale cloud, data cloud, scale data warehouse that we internally use for, all our analytics, you will use in your enterprise so that's one key thing the, second, key thing here is it's, highly secure. It. Allows you to do real-time streaming, directly into, the data warehouse so you can collect events, and at the same time ingest it directly into the warehouse and do, real-time analysis, somebody. Checks into the hotel and you want a analyst. To know that instantly, within under a second you can collect that put it into bigquery you know what it is and take actions on it so those are the kind of scenarios you can enable with this now, fully managed and support, standard sequel so most. Of your analysts would know some. Form of sequel and if they are if they can use sequel, for general analytics, bigquery. Is really, available available. For that one, of the things with bigquery we have done is we're. Making it really easy for you to get your advertising. Information. Data from different ad channels, and all especially with Google. Easily. Into the into, the into, bigquery for example, most of our customers. Already. Use Google tech, like Adwords double-click. YouTube. Play. Store all of these different things for advertising, all that, data with few clicks you can literally bullet into bigquery and so that's one thing and then if you have some of the data like we have lot of. Companies. That use drive. With, sheets and all to create like, some campaign information and stuff like that which, is like a master information, so if it is there and drive we, can just connect to it without even pulling it into bigquery so we do both, Federation, which is like connecting, to external sources but, also moving data into bigquery easily. So that you can you, can analyze, this at scale for, you other, than that we. Literally believe in democratizing, insights what, what, we do is if. You built like if one of the analysts has built. Segmentation. A query to go ahead and segment, your customers into, high. Value low value medium value based on let's say their lifetime spend you. Can go ahead and share that with every, other, analyst. Within the organization, using now. Using the query sharing capability. You can build a data, studio how, many of you have tried, or used data studio. Yeah. If you haven't you should absolutely, try, it out it's a it's, absolutely, free the, second thing is it has the latest thing that we announce the data studio is pre-built, gallery of reports, for, example if you're using AdWords, for advertising there's, pre build like. Report in. The report gallery that shows you all your ad spend analysis. And stuff like that but, one of the things you can do is just take that and connect with bigquery to start analyzing more, in depth what it is visualize.
It And share it with your with. Your customers, or within your organization. With, folks one, of the things people do a lot with data studio is they build those reports and actually share across organizations. So, you can take a report share it with other people they, can use their log in to see what is relevant to them and stuff like that so really powerful, tool sheets, integration, and and one of the most important things is we have huge partner ecosystem that also, allows you to get data into bigquery and. Analyze it and stuff like that so as I said earlier today. In. My introduction I was absolutely I was VP of engineering there for three and a half years and, built most of the data systems, before. That let me talk about Zulily Zulily is an online. Retailer a fashion. Retailer we. Launched 9,000 products every day they. Lasted on the site only for 72, hours so. There's a sense of urgency to, come and buy stuff because the deals go away it's a daily deal site. More. Than 5 million active, users, when. I left it's, a public number and. Then. What. Does that we used to have millions of people coming every day visiting. The site trying, to look for different offers and stuff like that. There, are few things which was really critical right when you have that, many people coming in that many products launching it's not a search based site it's basically a. Personalized. Discovery, kind of an experience that you give people you, have to show person, what is relevant at the right time at the right moment, and the, right price because if you don't do that there's. No way of searching, through the catalog which is so. Volatile, it's, not a fixed catalog system so, one of the things that we did which, was collect. All the data from our catalog, from all the different vendors and actually. Go ahead and personalize, the whole website based, on what, people were doing in past and leverage, machine learning for doing lot of that stuff. The, second thing was we. Had 500 merchants, or so like really large merchant, population, that was responsible, for building, they're bringing the right products in we, had to be able to give them right insights, in real time which.
Customers, Are buying, what, kind of products and how do you optimize that, in real time what should be the right prices, and stuff like that so that was there how, do you make. Better marketing, decisions, like the key thing was how, do you keep the growth on. Within. Six seven years we were like multi-billion, dollar like we. Company. And and that growth rate was really high how do you get more customers and how do you do better segmentation. Especially. Reducing, your cost of acquisition was. A key priority, for us so, here is what we did at a very high level it's a 10,000 foot, level thing the first thing we did was we. Realized when, we started on the journey we, need a single, source of truth for our customer, so. What we did was first, we went to bigquery we tried a lot of different technology options, on. Promises, option, we, tried open. Source stuff like hive Hadoop. Clusters and stuff like that what we realized was all of them were not fast enough we need an interactive. Level analytical. Capability, for our analysts, right they, are not going to sit there for 25 minutes 30 minutes fire a Hadoop, query and then come back and then get a coffee come back and it doesn't work like me, they were very clear like our, marketing analytics team was like no no we need interactivity. That is there otherwise it's not going to work so we leveraged bigquery, for all of that and bigquery we created. Some. Key data sets there's, one data set which was all about customer, so, we had all the customers, or members that, we had members our customers. Who haven't bought anything yet everybody. Had a row we, had all the information for them within a single row and it, was like how many times have they bought what have they bought what's total, money all that was in a single row that we could get all their, activity, was in a related, table so, that means you had user here's all the activity, on the product and always dimensional, so it was already there and then, we created so. That was one key thing like everything. About your customer we'd there's, a lot of talk about CDP's. Now customer data platforms, and stuff like that it's fundamentally, just a central place with all your customer activity information, once, you have that now you can do segmentation you can say hey tell me all the customers, that, have come in last seven days multiple, times tell, me all the customers that came seven, multiple. Times last week and haven't come back again so we were able to create these dynamics, segments, in bigquery, at scale and we. Means like the analyst team in marketing, analyst team could start creating those and start, segmenting them because, previously what we we had was hey we have lot of customers, and we send them emails and that is it like everybody. Gets an email and and stuff like that but this allowed us to do micro segmentation at the scale we could never do and so, that way then those analysts, were able to make those decisions segment.
Them Figure out anybody who hadn't come last week should. We send them an email and remind them or maybe give them an offer or if. Somebody, has been coming consistently, not buying it. Should we show different products so those kind of decisions were easier to make at scale, we. Had like, when. I left it was more than a petabyte scale data. Total, data set that we had but we. Started collecting a lot of information now the second thing as I said we basically collected, real-time events from the site way as soon as people landed on the website all, the activity. They did what, products they were clicking on what products they were buying what products were going in the cart we collected, all of that information the. Statistics, which is interesting it's actually in the case study. When. We started, the project like for. Real-time collection, we, are collecting 50 million events in a in. A day by. The time I left, we were already at five billion events a day so, it was an 100x, increase, in number of events the thing is once, you start seeing value from data you want more and more collection, and the, data volume just grows so. We went from 50 million to five billion a day we're, peaking at like almost fifty thousand or thirty five to fifty thousand events a second but, the key thing there was getting. That information in real time ability, to aggregate that information push. It back into the serving platform, like website to, go ahead and see how many people are watching some products are making, decisions like that became really easy and and we didn't have to worry about anything because the scaling of the infrastructure, was all taken care of by by. The cloud. Infrastructure, that we had so. We basically, I put all that data leverage. Data frog for the cleansing, of data making sure the data was ready for analysts and then all the data was put into bigquery. Analysts. Had tableau. So they use tableau for their, visualizations. Reporting. Some, amazing reports to see trend lines about how, many people were coming last year same day this, day today, what has the change is it going up down how, many people are ready to churn out because we had a metric on active, customers. All of that cuz the analyst teams could do by. Themselves and then of. Course we also took. The same data the single view of customer and their. Activity, and all of that and also powered our operation systems we had more than 70 api's. That were built on top of our, data sets within bigquery, that, allowed different systems to call in and say hey give me the, performance, of the supplier so we can give a better offer better. Estimation. For our customers, how long it takes to ship them and all that so all those kind of things were able to do just because how. We pick. The infrastructure. Leveraging. Bigquery for a lot of those those. Kind of scenarios one. Of the things we launched this this, week was bigquery ml it's, a machine learning capability, directly inside bigquery, what. It is is very simple you have marketing, analysts, they. Know bigger sequel, they've been running, sequel, queries, analyzing. Stuff, bigquery, ml, actually, allows you to do regression, models. Like linear regression, logistical, regression directly. Inside the data warehouse so, if you have pulled all your data moved it into a data warehouse, which is already there now, you can write two lines of code to create regression, models and do, things. It's. Pretty interesting and for prediction you just have to say select ml dot product and you get predictions, back so it's as simple as that now. For the experience, from from, an analyst perspective, here, are a few examples you, must have if you saw how many of you saw our day two keynotes. Migrated, you heard from, 20th. Century Fox how they were able to go ahead and do a linear regression to go ahead and predict which customers, were going to come back for Maze, Runner compared. To all the things that they have done in past in different movies, and also that's one example there. Were a couple of other sessions we were talking about but. Like simple, things like you. Want to go ahead and do logistical, regression for classification which. Of your customers, are high-value, which ones of them are going to churn which, ones are going to be. Doing. The subscription, or not taking, the subscription, all that, kind of analysis, now you can really do it inside and run, a whole machine learning model within seconds, and do it one. Of the other things is we also have our partners like looker and all building. Experiences, to bring all of this together now, within it so that's the bigquery ml we just launched this this. Week you should try, it out see how you can leverage, that for different marketing scenarios. That. You have, yeah. Other, than that we have a huge portfolio. For your AI and m/l capabilities, right from, auto ml capability, for translation.
That We launched a natural. Language as. Well as auto ml for vision which was already there those, are just great, tools for, non-technical. Users. Or developers to, just move data in tagged. Them and then start using the API is or batch predictions, to board and figure out what and what you want to do additionally. If you have data scientists, and PhDs, within your organization, they can do complex, neural, network based. Modeling. You, can absolutely use, ml. Engine tensorflow, all of those things to go ahead and run, those kind of models the, other key thing we also have is we have this thing called advanced. Solution lab where we work with key customers and partners to work together on solving some. Interesting. And major problems, that you may be facing in. In partnership, so that's there for. The next thing I want to touch upon is open source solutions. Focusing. On open source is key value. Proposition, for us we want to make sure all, the stuff that our customers, are investing, in is it's. Possible for them to leverage them not just on our cloud but also on. The different environments you have and then. Containerization. Everybody, who's. Been using kubernetes, or has. Looked at it it's basically one. Of the fastest-growing open source projects, started, from within Google we were using containers, making. It available. I'm. Not going to go in depth of this but in 15 years of, innovation, in Big Data we have been before. Cloud, was there we used to actually publish all these papers and share whatever, we were learning now we are making them available, as, different products like bigquery spanner. Lot of other products or we also have products that provide, managed, open, source solutions, like data. Proc if you have a Hadoop, environment. How. Many people have Hadoop environments, within their Hadoop. SPARC stuff good, so. Data proc is basically managed Hadoop and spark in the cloud so. If you you, don't have to worry you don't need administrators. You don't need management capability, just write your scripts, in hive, or, SPARC and just run it and we, just take care of the whole loading. Of the whole machines, bringing, it down scaling. Up and down auto scaling those are all the investments that we're making from our side and then, composer is airflow which is if. You need a distributed. Orchestration. Capability. Pull. Data from point, A go, ahead and do segmentation after. Segmentation, move it to Adwords. Or our Facebook, or gdn and then, go do something else and send an email that kind of orchestration, you can do in in airflow, pretty easily yeah, other, than that you. Saw some of you must have seen SST, or demo yesterday pretty, cool stuff that's. What we are doing from an open source perspective is to your kubernetes, tensorflow. Again for. Machine learning is all all open, source from our side as. I said earlier one of the key differentiators, we, have is our global presence and reach. Basically. If you look at our global network this, is the network that is powering all of our the, billions of users that we go ahead and, like. You know provide solutions, for is the same network, we are making available for our customers, it's. Basically. You. Can locate services closer to your user. Imagine. YouTube, being watched in all the different countries how, close we need to be to those users that's the same network that that. We leverage for GCP, it's. Highly available and then of course now. Choose your resources and locations where you are like if you want to go ahead and regionalize. Surf because of data locality and all maybe, provide that now. If you look at our infrastructure. And reach we. Have hundreds of thousands of miles of cables, fiber-optic. Cables we, actually have submarines, that actually put down all these cables through oceans and stuff like that so it's pretty cool now. There are we now are in, fifteen regions, that, under color six. More coming in 2018, but we had 15 regions 55, zones already, so, imagine the whole scale that we that we are putting together now with that I would like to call upon. My next guest to, talk more. So. Yeah. We are build scalable attack. System to handle global, traffic within. Two years or things, to Google, infrastructure.
So. What, does the scalability. Mean to start up like us, we. Started as our two co-founders thorough. Three. Years ago and. For. The first CEO actually moved seven times and. Now. We started, this year with the 20 team, members now, we are actually reaching 250 team members and we, are supporting enterprise, customers, with a tens more than like 10,000 employees right. So, throughout. This like our journey we always, use, our infrastructure. On top of Google cloud so, no, matter we have two team members 50. Team members or, supporting, 10,000, employees Google. Infrastructure, was, able to handle the scalability. So. We're over overview, Obama local we. Are over, at a company, and our, vision is atoning, your first party data into. The add performance you. Know secure. You, know a. Safe. Way to give. You full control of over your data for. Last three years are we, on board more than like 60 clients our goal, for the last two years is, testing. Our machine learning algorithm, on different. Countries and different verticals, so, we are now running our campaigns, on 40 different countries you, know for the call from mobile, commerce rather, sharing, food delivery to. All. Different, kind of own on-demand, apps. So. Whenever. We meet new clients we, are very excited, about talking to new clients a, lot. Overcome we a lot of startups turning, their idea, into, directory services, and then, don't. Worry about the, service care anymore. Because all they are all backed by cloud computing, right but. When we meet when we talk to companies, where, they actually are stock. Ad is a lot, of company has pain for the growth and for. The monetization. Need and the reason why is we'll cover this in the next slide but, complexity. To build mobile. Marketing. And. Monetization. Platform is becoming, more complex. So. Google. As we covered in the previous slide. Our. Goal our goal and Google's goal is to support you you need you, know in a fast way so, your growth and, business. Goals cannot be slowed down so, our goal, is, delivering. Best. Modern, technology. Within, a quorum so. This. Is what we did for last two years are we, now, handle, more than 600. K7q. PS this. Is a sixty billion daily. Requests. Per day, so. What. Was sixty billion requests. Mean is, let's. Say you, build some messenger services. Without. 100 million people if they text, like six six hundred times text per day there's, a 60, billion per day so, this system is built basically. In a scale over something. Like a large-scale messenger system in a country so. We manage this carnival, scale of services, with a nine infrastructure. Team members, and. On. Our, data pipelines, we run 200. Different data flow graphs for on behalf of our customers on, different verticals, and. We, handle more than hundred terabytes data. Today with, the forty thousand virtual CPUs. So. The complex so while we build this system we, realize, the complexity, over. Mode, on attack, and marketing text system so, in all the desktop, earth right, what. Does our ad serving system means it's basically a content delivery system you can put your image, on. On, your content delivery services. And. You can put a simple account row on your web pages you, can account impression, and, if, your client is we can just count clicks but. Now actually your, marketing. Department or your, advertiser, they want optimize for the post clip post installation events, which is happening on, advertisers. Domain, or advertisers, app so. Now I don't lecture you just need to serve that but, also you need to ingest, the, data from, your advertiser. Your clients, app so. Now you need a, not. Only the DSP services. But also data in JSON services, in the meantime you got to analyze a LTV customer, acquisition cost then, is a lot of big data system, which is a TMP, and, also. You need to have a you, need to support internal, analytics groups, so which is neither analytics. All the players right so, now, we are in the third day of GCP you probably heard about ton of different, data.
Infrastructure, From spinner cloud. Storage. MEMC. Called. BigTable. Right, so, what, should you use right that's what we kind of tried to figure out the last two years so, each module each, icon here, is actually turning into the proper. Google. Cloud infrastructure, for, example on, our serving infrastructure we, use current, ease and the, GCE and for, the TMP we, actually realized, that BigTable, is most, fastest, and scalable, way to support, large-scale and infrastructure, for the analytics, read big quarry and for. The big data pipeline yes we tried several different solution we realized, that data flow is the best one. And. At the last are the, products. The, brain. Part right. Smarten, is how to optimize, for to, add performance how to deliver the, better revenue and better cost, optimization. For. Your marketing, needs is a key motion only and we, decide, to use tensorflow we recover, born this one in the our last slide so. Yeah. This actually shows that as complexity, or building, modern, architecture. And. We. Actually, will we actually figure it out although, what. Is the best solution for this one with Google and now, that you can use, this system to, launch, your, ad plan marketing, platform the Covidien three months. Another. Interesting, observation we, realized. When you talk to our customer, is, consolidation. Over marketing. World. In global, menno so. Previously. Like when. You run advertising, campaign what does it mean you normally call agencies, right, and hey, can you put our ads on next, to 101 or this newspaper, so let's say you go to different country, they say you are expanding. To Japan, you normally talk to some, big, agencies, like, a tenchu or some traditional agencies, but. What we find very interesting is now actually the top tech, companies, they, are consolidating. There, are ad planning. Team and ad, execution. Team in one. Location or all only. Very handful like regional, areas, so for example one of the largest streaming, video, streaming company they run all advertising. Campaign from los carros or one. Of the largest. Like a home sharing company they run all the campaign from, San Francisco and Singapore, foot and to cover APEC, and global. The reason why is, marketing. Is not only are not just about the, page. Views or publication. It's all about the big data this. It becomes very hard to duplicate, your big data team, for every, country right, that's a why actually are consolidating, their, marketing, big data team and the execution, team in new of your countries so now actually given, that your, infrastructure, serving, system got to support global traffic. So. The, question is now, some, of the ad exchanges, right for example we are also not only part of the Google cloud we are also Platinum Google EDX exchanges. They, have, a multiple. Ad, exchange. Servers, across, the world cover the global traffic some, of the smaller, ad exchanges, they, only have our services services, in. U.s.. Some. Of the local exchanges, they have a server in Japan, in Singapore, right to. Figure it out where they are and, to connect and to design. You infrastructure. For, this country from, the, global, set up is also, time wasting and time cause it is not rocket science should, not be focused, over your business but, but it's easily waste you're like six months or one you know right now, we, don't know on, the website you also got to interior, with uh all these mobile, measurement, companies I mean in, previous, like that abilities comScore nowadays as flyer tune some.
Of Them have obvious servers all in us some, of them have in Japan some of them have a multiple, location in global, location again, to, figure it out we, don't want wait, we don't want to like waste time on this kind of all under works so, we actually figured out everything and Google. Infrastructure, support, or just global, scales for. Example we run, our BigTable. Instances, in four, different locations in, in, the world and we. Run our machine, on intense flow engine on us. Only so, we actually run, some, of the dictator, pipelines through the data flow you know for different locations we. Actually send intermediate, data to. Save, the the. Network, traffic cost between the data centers and the run all the machine learning jobs from. The US and then this will the result. Of the machine learning reject to the, other location in. The. Last like the most like exciting, and the important part is a machine learning how, to turn your data, into, the real performance right so, as. A studio covered in the previous session Google. The benefit is you can cover end to end, what does end it end means right so, you have a great services. Your, service to local data and you, basically, to low but it has a feedback from your users right, now you've, got to turn that data into the, actionable items through the machine learning and then. The action of data the actionable item should be reflected, to your services, now, your, users, react to the new, system and then, they'll give another, PDA so this, kind of feedback loop so your business success is depends, on how, you firstly, iterate this system and. We, built all this entertain it. Right iteration, feedback system so on top of a Google data infrastructure, so, for the robot collections. User. Profile stories we use the pops up BigTable and then, we use data flow to in. Extracted. Data, from. Those data, pipeline and we, feed that data into, the pencil flow, so. When. You first begin. Yes. It's two years ago tense flow was of it only and then we also small so we started, up from our, sidekick. Like a traditional, machine learning sitting right and now as, everyone, knows you, got to switch to the deep learning system, right so the, peripheral of a tense flow and the Google infrastructure, is tends, to flow a lot of people think all is image recognition oh it's deep learning it's not it's it also support all the traditional, models so, we used to run different models on a different machine on infrastructure, now with the tensorflow we, actually run all the traditional, leader or logistic. Regression models and also keep learning news team learning models and sometimes we mix those two models hybrid, models right we test all different kind of models on top of tensorflow and, we. Compare. The result of these different models on a, through. The a B test platform, and. We realized, in different vertical, in different, or data. Channel. Sometimes. It's even a traditional model works better so still our serving, system can support all these different kind of solving model for different clients need without, like any other, extra infrastructure, cost and on. Top of total. Variance. Of these models right most, exciting part in not only the motor for, us on the line physical, layer so. We started from the CPU based training. We. Are using GPU, and most. Exciting, stuff. Is now if Google is are launching the, specialized. Processor, for the tensor, flow rate tensor flow unit or GPU so. We can test all these different kind of. Physical. Infrastructure, without, in changing, any of our infrastructure. And. Yeah. Result or we. Achieve, up. To like three times better cost, savings or.
Revenue. From. For user base in many, of our clients. So. To wrap up. We. Grew. Tremendously. For, the last few years without. Worrying any, of the server infrastructure, or, spending. A lot of money for, the DevOps team, we. Hope you can grow with, a Google like, that like. Us and. We. Are very excited about the new and trend watching learning infrastructure. So, what. We like, about Google is. Google's. Vision on the scalability, and also. On, the. Global. But on top of that we, what we actually really like about Google is their, philosophy on, the openness right it. Actually really resonate to our philosophy on, keeping. Your data on the full control, on you. And. How. To value. Your first party data using. This kind of new technology that's, actually our what, we our vision is right, thank you. You.