The next edition of the HP Discover Podcast Series delivers an innovation case study interview that highlights how data-intensive credit- and debit-card marketing services provider, Cardlytics, delivers millions of highly tailored marketing offers to banking consumers across the United States.
Cardlytics, in adopting a new analytics platform, gained huge data analysis capacity, vastly reduced query times, and swiftly met customer demands at massive scale.
To learn how, we sat down with Craig Snodgrass, Senior Vice President for Analytics and Product at Cardlytics Inc., based in Atlanta. The discussion, which took place at the recent HP Vertica Big Data Conference in Boston, is moderated by me, Dana Gardner, Principal Analyst at Interarbor Solutions. [Disclosure: HP is a sponsor of BriefingsDirect podcasts.]
Here are some excerpts:
Gardner: At some point, you must have had a data infrastructure or legacy setup that wasn’t meeting your requirements. Tell us a little bit about the journey that you’ve been on gaining better analytic results for your business.
Snodgrass: As with any other company, our data was growing and growing and growing. Also growing at the same time was the number of advertisers that we were working with. Since our advertisers spanned multiple categories — they range from automotive, to retail, to restaurants, to quick-serve — the types of questions they were asking were different.
So we had this intersection of more data and different questions happening at a vertical level. Using our existing platform, we just couldn’t answer those questions in a timely manner, and we couldn’t iterate around being able to give our advertisers even more insights, because it was just taking too long.
First, we weren’t able to even get answers. Then, when there was the back-and-forth of wanting to understand more or get more insight it just ended up taking longer-and-longer. So at the end of the day, it came down to multiple and unstructured questions, and we just couldn’t get our old systems to respond fast enough.
Gardner: Who are your customers, and what do you do for them?
Growing the business
Snodgrass: Our customers are essentially anybody who wants to grow their business. That’s probably a common answer, but they are advertisers. They’re folks who are used to traditional media, where when they do a TV or radio ad. They’re hitting everybody, people that were going to come to their store anyways and people who probably weren’t going to come to their store.
We’re able to target who they want to bring into their store through looking at both debit-card and credit-card purchase data, all in an anonymized manner. We’re able to look at past spending behavior, and say, based on those spending behaviors, that these are the types of customers that are most likely to come to your store and more importantly, most likely to be a long-term customer for you.
We can target those, we can deliver the advertising in the form of a reward, meaning the customer actually gets something for the advertising experience. We deliver that through their bank.
The bank is able to do this for their customers as well. The reward comes from the bank, and the advertiser gets a new channel to go bring in business. Then, we can track for them over time what their return on ad-spend is. That’s not an advantage they’ve had before with the traditional advertising they’ve been doing.
Gardner: So it sounds like a win, win, win. As a consumer, I’m going to get offers that are something more than a blanket. It’s going to be something targeted to me as the bank that’s providing the credit card. They’re going to get loyalty by having a rewards effort that works. Then, of course, those people selling goods and services have a new way of reaching and marketing those goods and services in a way they can measure.
Snodgrass: Yeah, and back to this idea of the multiple verticals. It works inside of retail, just as well as restaurants, subscriptions, and the other categories that are out there as well. So it’s not just a one-category type reward.
A customer will know quickly when something is not relevant. If you bring in a customer for whom it may not be relevant or they weren’t the right customer, they’re not going to return.
The advertiser isn’t going to get their return on ad-spend. So it’s actually in both our interests to make sure we choose the right customers, because we want to get that return on ad-spend for the advertisers as well.
Gardner: Craig, what sort of volume of data are we talking about here?
Snodgrass: We’re doing roughly 10 terabytes a year. From a volume standpoint, it’s a combination of not just the number of transactions we’re bringing in, but the number of requests, queries, and answers that we’re having to go against it. That intersection of growth in volume and growth in questions is happening at the same time.
For us right now, our data is structured. I know a lot of companies are working on the unstructured piece. We’re in a world where in the payment systems and banking systems, the data is relatively structured and that’s what we get, which is great. Our questions are unstructured. They’re everywhere from corporate real estate types of questions, to loyalty, to just random questions that they’ve never known before.
One key thing that we can do for advertisers is, at a minimum, answer two large questions. What is my market share in an area? Typically, advertisers only know when customers come into their store with that transaction. They don’t know where that customer goes and, obviously, they don’t know when people don’t come into their store.
We have that full 360-degree view of what happens at the customer level, so we can answer, for a geographic area or whatever area that an advertiser wants, what is their market share and how is their market share trending week-to-week.
The other piece is that when we do targeting, there could be somebody that visits a location three times over a certain time period. You don’t know if they’re somebody who shops the category 30 times or if they only shop them three times. We can actually answer share-of-wallet for a customer, and you can use that in targeting, designing your campaigns, and more importantly, in analysis. What’s going on with these customers?
Gardner: So the better job you do, the more queries will be generated.
Snodgrass: It’s a self-fulfilling prophesy. For us, with Vertica, one of the key components isn’t just the speed, but how quick we can scale if the number of queries goes up. It’s relatively easy to predict what our growth and data volume is going to be. It is not easy for me to predict what the growth in queries is going to be. Again, as advertisers understand what types of questions we can answer, it’s unfortunately a ratio of 10 to 1. Once they understand something, there are 10 other questions that come out of it.
We can quickly add nodes and scalability to manage the increase in volumes of queries, and it’s cheap. This is not expensive hardware that you have to put in. That is one of the main decision points we had. Most people understand HP Vertica on the speed piece, but that and the quick scalability of the infrastructure were critical for us.
Gardner: Just as your marketing customers want to be able to predict their spend and the return on investment (ROI) from it, do you sense that you can predict and appreciate, when you scale with HP Vertica what your costs will be? Is there a big question mark or do you have a sense of, I do this and I have to pay that?
Snodgrass: It is the “I do this and I’ll have to pay that,” the linearness. For those who understand Vertica, that’s a bit of a pun, but the linear relationship is that if we need to scale, all we need to do is this. It’s very easy to forecast. I may not know the date for when I need to add something, but I definitely know what the cost will be when we need to add it.
Compare and contrast
Gardner: How do you measure, in addition to that predictability of cost, your benefits? Are there any speeds and feeds that you can share that compare and contrast and might help us better understand how well this works?
Snodgrass: There are two numbers. During the POC phase, we had a set of 10 to 15 different queries that we used as a baseline. We saw anywhere from 500x to 1,000x or 1,500x speed in return of getting that data. So that’s the first bullet point.
The second is that there were queries that we just couldn’t get to finish. At some point, when you let it go long enough, you just don’t know if it is going to converge. With Vertica, we haven’t hit that limit yet.
Vertica has also allowed to have varying degrees of analysts’ capabilities when it comes to SQL writing. Some are elegant and they write fantastic, very efficient queries. Others are still learning the best way to go put the queries together. They will still always return with Vertica. In the legacy world prior to Vertica, those are the ones that just wouldn’t return.
I don’t know the exact number for how much more productive they are, but the fact that their queries are always returning, and returning in a timely manner, obviously has dramatically increased their productivity. So it’s a hard one to measure, but forget how fast the queries have returned, the productivity of our analyst has gone up dramatically.
Gardner: What could an analytics platform do better for you? What would you like to see coming down the pipeline in terms of features, function, and performance?
Snodgrass: If you could do something in SQL, Vertica is fantastic. We’d like more integration with R, more integration with software as a service (SaaS), more integration with these sophisticated tools. If you get all the data into their systems, maybe they can manipulate it in a certain way, but then, you are managing two systems.
Vertica is working on a little bit better integration with R through distributed R, but there’s also SaaS as well. In a SaaS shop, there are a lot of things that you’re going to do in SaaS that you are not going to go do in SQL. That next level of analytics integration is where we would love to go see the product go.
Gardner: Do you expect that there will be different types of data and information that you could bring to bear on this? Perhaps some sort of camera, sensor of some sort, point-of-sale information, or mobile and geospatial information that could be brought to bear? How important is it for you to have a platform that can accommodate seemingly almost any number of different information types and formats?
Snodgrass: The best way to answer that one is that we don’t ever want to tell business development that the reason they can’t pursue a path is because we don’t have a platform that can support that.
Today, I don’t know where the future holds from these different paths, but there are so many different paths we can go down. It’s not just the Vertica component, but the HP HAVEn components and the fact that they can integrate with a lot of the unstructured, I think they call it “the human data versus the machine data.”
It’s having the human data pathway open to us. We don’t want to be the limiting factor for why somebody would want to do something. That’s another bullet point for HP Vertica in our camp. If a business model comes out, we can support it.
You may also be interested in:
- Efficient big data capabilities help Cerner drive needed improvements into healthcare outcomes
- Learn how Visible Measures tracks an expanding universe of video and viewer use big data
- Using Vertica, Empirix delivers complex carrier network performance data
- With big data, the DNC turns politics into political science
- Need for quality and speed powers Sentara’s applications modernization journey
- Big data changes the customer analysis game for Yammer, Spil Games, Jobrapido
- Application development efficiencies drive Agile payoffs for healthcare tech provider TriZetto
- How MZI HealthCare identifies big data patient productivity gems using HP Vertica
- Advanced IT monitoring Delivers Predictive Diagnostics Focus to United Airlines
- HP Vertica Architecture Gives Massive Performance Boost to Toughest BI Queries for Infinity Insurance
- HP-Fueled Application Delivery Transformation Pays Ongoing Dividends for McKesson
- Podcast recap: HP Experts analyze and explain the HAVEn big data news from HP Discover