This case study panel discussion highlights how various organisations are coming up with the means to develop far better analytics about their customers. Learn how high-performing and cost-effective big data processing enable a steep learning curve from customers on their wants and preferences.
The expert panel consists of Rob Winters, director of Reporting and Analytics at Spil Games, based in Amsterdam; Davide Conforti, Business Intelligence director at Jobrapido, based in Milan; and Pete Fishman, director of Analytics at Yammer in San Francisco.
Below are some excerpts.
Businesses have been analyzing their customers for a long time. What's different now?
Fishman: We're a cloud software service, and the data is big. Our data on the customers is now all living in a central place. By aggregating across companies that are using your software, you can get really significant sample sizes and real inference, both from an economic sense, in terms of measuring the lift, but actually, because the sample sizes are so big, you can get statistical inference.
That's the starting point for making analytics valuable and learning about your customers.
Winters: For me, the problem space is extremely different from what I was dealing with a couple of years back.
I was in telecom before this. There, you're dealing with 25 million people, and if you rescore them once a month, that's fast enough. On a web scale problem, I'm dealing with 200 million customers and I have to rescore them within 10 or 15 minutes. So you're capturing significantly more data. We're looking at billions of records per day coming into our systems. We have to use it as fast as possible, because with the customer experience online, minutes matter.
Conforti: It's absolutely the same story with us. We have about 40 million unique visitors per month now. We've grown by double-digits since our start as a startup in 2006. Now, everything is about user interaction, how our users behave on-site, and how we can engage them more on-site and provide them a tremendous ad-hoc user experiences.
Winters: We're primarily a platform. We do some game development and publishing, but our core business is just being the platform where people can come and find content that's interesting to them. We've been around for about nine years.
We started out as just a Dutch [gaming] company and then we've acquired other local domain names in a variety of languages. At this point, we have about 50 different platforms, running in about 20 different languages. So we support customers from all over the world. In a given month, we have over 200 countries with traffic onto our sites.
The entire business is changing, and you're competing based off that customer experience that you can deliver. We have a couple target audiences: girls, young girls, 8-14; boys; and then women.
Fishman: Yammer is a startup in San Francisco. We were acquired about a year ago by Microsoft and we're part of the larger Office organization. We view ourselves as enterprise social, taking this many-to-many communication model and making communication at your company much more efficient.
It's about surfacing relevant knowledge and experts and making work lives better. I run an analytics team there, and we essentially look at the aggregate customer behaviors and what parts of our tool people are using.
This was a really revolutionary idea that our founders David Sacks and Adam Pisoni had, way back when Facebook wasn't nearly as relevant as it is today. But we've leveraged a lot of the way that people have learned to interact in their social life and bring some of that efficiency of communication. They saw that these social networks would grow and be relevant in a private, secured context of your business.
Conforti: Jobrapido started in 2006 as an entrepreneurial challenge that Vito Lomele, an Italian guy, started in Milan. It's quite a challenge to live in the online market in Italy, because talent pooling isn't as wide as in U.S. or in other countries in Europe. What we do is provide job-seekers the opportunity to find their new job.
We're an online job-search engine and we currently operate in 58 different countries with more than 20 languages. We're all in this big headquarters in Milan with a lot of different nationalities, because of course, we provide the service in local languages for most of our customers.
Recently, we have been purchased by the Daily Mail group, a big media group based in London. For us, it's everything from job-seeker acquisition and retention and engagement deals with constant quality and user experience on-site. We use our big data warehouse in order to understand how to better attract and retain customers on the basis of their preferences. And we also use it to tweak our matching algorithm, which works more or less like a Google algorithm.
We crawl a lot of contents from different sources, both job boards and other job sites or directly in the working pages of individual companies. We put them together in a big database and, using statistical tools, we infer which kind of rankings our job-seekers are willing to see.
So it's a pretty heavy data crunching exercise that we do everyday on millions and millions of different sponsored or organic postings.
For example, if Yammer guys or if Spil Games guys want to hire a software engineer, they can directly promote their sponsored ads on Jobrapido without having to sponsor them on a job board. So we're trying to aggregate and simplify the chain of job search.
What was the problem you had to solve when it comes to getting at this big data for analysis?
Winters: For me, the challenge was multi-fold. How do you deal with this data problem, with this variety and volume information? How do you present it in a meaningful fashion for employees who've never looked at data before, so that they can make good decisions on it? And how do you run models against it and feed that back into a production environment as quickly as possible, so that you can give those customers a better experience than they were ever getting before on your platform?
My problem was that no one had ever tried to do it in my company before. We walked in with effectively a clean slate. But as you start to bring in different data sources, you start with all the stuff that you know you're going to need right away.
You start seeing needed links for other data sources. At this point, we're pulling data from thousands of databases, merging with dozens of application programming interfaces (APIs). You're pulling in your web log data, so that you can personalize for those folks who aren't giving you registration information.
When we first started looking for a data warehouse appliance or application, we were running Postgres with no indices, just copies of production data. For data guys, that means that a query will take eight hours to execute. It's a table of a couple of million rows.
We knew that a typical row-based solution was out. So we started looking at some of the other applications out there. The big ones are Teradata, Exadata, and Greenplum, but you're going to have to mortgage the house of every employee in the company to be able to afford a license for those applications, and we're a pretty small company. So those were out.
Then, we started looking at some of the other boutique vendors like Infobright, and basically we saw that with HP Vertica, we can have relatively low load on our database administrator (DBA), so we can develop quickly without a lot of maintenance.
The pricing model fits what we need to achieve, and the performance is so good that we don't have to spend a ton of time on optimization now. We can basically move very rapidly along this path of becoming a data-driven organization without having to get held up on index optimization or trying to optimize our queries and rewrite paths.
We can just throw a lot of stuff into the system, smash it together, take the results, and get big wins for the company quickly.
We have a data center, and we do everything on our own private servers. For us, the next step is probably going to be moving more into a private-cloud model, and hopefully, Vertica will work in that environment as well.
At Yammer, what was your big data problem and how did you solve it?
Fishman: Our problem set was that there were a lot of people trying to get into the enterprise social space. A lot of social networks are popping up, and essentially competing for attention at work is a challenge.
We felt that data was necessary to have a competitive advantage. David Sacks and Adam Pisoni had a vision of developing a consumer software company with rapid iteration. With that rapid iteration you get an extra advantage if you're able to reorient yourself based on what part of the product is working. Our data problems were largely about making data be a competitive advantage in our development methodology.
It's about speed
What was it about Vertica that was instrumental to the point where you've adopted it? Is it a concurrency issue, a volume issue, speed, or all the above?
Fishman: It's all of the above, but the real highlight is always going to be about speed, especially, given the incredible competition for talent, not just in the Bay Area, but all over, especially in the data field.
Anybody that has data in their title is someone that's highly sought after. That ability to minimize the cycle times for those folks who are such a challenge to keep and get excited about the projects that they're working on and is a tremendous solution that allows them to maximize their own abilities is really critical. It's the same in our space, and in software development in general.
When we take on these big risks and challenges, the ability to very quickly identify whether we're going in the right direction, and then reorienting where we are going, has been really critical to Yammer being successful.
Davide, how did you get a handle on data problems?
Conforti: When I joined Jobrapido, we already ran tons of A/B tests, which are the lifeblood of our product innovation. We want to test everything, from changing the color or the font of one button to a different layout, because these have tremendous impact on improving the user engagement.
Before, we used the Google Analytics tools, but we didn't like that much, because it's sample data, so you hardly reach statistically meaningful results. We decided to build a data warehouse to assure flexibility, performance, and also a higher level of control and data consistency. That's end-to-end control from the source, toward the visualization, in order to make them more actionable in terms of product development.
With Vertica, we did exactly this. We poured all the different data sources into one bucket, organized it, and now we have a full control over the data model. With my team, I manage these data models. It's fascinating how fast you can add pieces to the puzzle or remove others that are no longer interesting, because our business model, of course, is a living animal, a living creature.
We really appreciate this flexibility and the high level of control that Vertica allows. This improved a lot our innovation throughput and it's going to improve it even more in the future.
Currently, we crunch on Vertica about 30 GB of data everyday (i.e. we upload 30 GB/day on Vertica). But we're going to double it in a few months, because we're adding more stuff. We want to know more about the click patterns of our job-seekers on the site, and this is massive data flowing into Vertica. Also, our licensing in terabytes will likely double in the future.
Another hard fact that I can share with you guys is that every one of you using Vertica doesn't have to be satisfied with the first implementation of the query. If you're able to optimize it, you almost increase the performance of the query by more than 100 percent. This is my personal experience with consultants or advisers. Vertica is happy to provide the support, and this is really value-adding.
Winters: As far as metrics of success, when we were doing our proof of concept (POC), we looked at primarily query performance. At that point, we weren't looking at using it for prediction and personalization, but just for analytics and reporting.
What we saw was against an indexed Postgres database. We had done some optimization on the data. Our queries were running more than 1,000 percent faster, and Vertica was scaling pretty linearly, whereas with Postgres, when we put more data into the tables, they just started choking and just died completely.
For me, it allowed me to actually do my job and have my team do their jobs, which is a pretty big metric of success.
The other thing is that with a relatively small cluster, we can support hundreds of people and reports directly accessing the database, a dozen analysts or people who directly query information out of the database, and all of our personalization activities simultaneously with minimal performance hiccups. That's a big metric of success.
Fishman: I have similar feedback as Rob, which is a comparing against a Postgres database. The speeds are at least one -- and probably closer to two or better -- order of magnitude faster. Certainly on the cost side, it's important with data to consider the whole cost. So this is sort of a theme.
There is a cost in a variety of managing and teasing out the useful insights that aren't necessarily in the sticker price. When considering a data solution, people should consider the end-to-end costs. What's really the cost per insight, as opposed to the cost per terabyte or the cost per whatever.
We certainly feel that Vertica has been our best solution. We've been customers for over three years. So it's quite a long relationship. I couldn't imagine going back to a multi-day query, or something like that.
One thing that Davide mentioned is that he's forecasting how much data he will be putting into Vertica. I'm a forecaster myself by trade. Back in 2010, we were doing some estimates of where we would be by the end of 2011 in terms of our data volumes. This is a pretty simple extrapolation, and I got it wrong by at least an order of magnitude.
What we found is that when you start to get real insights from data, you want to get a little bit more, collect it maybe here or there. Also, as our product was growing, we faced some real exponential growth on the data and adopted clever solutions for maximizing that metric that we care about -- cost per insight, or minimizing the cost for insight.
There are many things going on simultaneously. So tripping over really valuable insights can happen a lot more easily than when you're more naïve about it. Essentially, you're facing headwinds in that. Finding insights become harder. At the same time, you have larger data volumes and some economies of scale there. So there are a lot of things simultaneously interacting, but clearly one thing to drive down that metric is best-in-breed tools.
Of course, best to get the information of the people who can use it than to simply look to cut cost.
Fishman: Of course. If you view analytics as a cost center, that's the wrong view. It should be aimed at optimizing revenue streams. We micro-optimize the product, we micro-optimize sales and marketing, the business. Analytics is about improving everybody at their job, making data available to allow people to be more effective.