Visible Measures measures via a massive analytics capability an ocean of video at some of the highest scales I've ever heard of. By creating very deep census data of everything that's happened in the video space, Visible Measures uses unique statistical processes to figure out exactly what patterns emerge within video usage at high speed and massive scale and granularity.
To learn more about how Visible Measures measures, please welcome Chris Meisl, Chief Technology Officer at Visible Measures Corp., based in Boston.
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: Tell us a little bit about video metrics. It seems that this is pretty straightforward, isn't it? You just measure the number of downloads and you know how many people are watching a video—or is there more to it?
Meisl: You'd think it would be that straightforward. Video is probably the fastest growing component of the Internet right now. Video consumption is accelerating unbelievably. When you measure a video, not only you are looking at did someone view the video but how far they are into the video. Did they rewind it, stop it, or replay certain parts? What happened at the end? Did they share it?
There are all kinds of events that can happen around a video. It's not like in the display advertising business, where you have an impression and you have a click. With video, you have all kinds of interactions that happen.
You can really measure engagement in terms of how much people have actually watched the video, and how they've interacted with a video while it's playing.
Gardner: This is an additional level of insight beyond what happened traditionally with television, where you need a Nielsen box or some other crude, if I could use that term, way of measuring. This is much more granular and precise.
Meisl: Exactly. The cable industry tried to do this on various occasions with various set-up boxes that would "phone home" with various information. But for the most part, like Nielsen, it's panel-based. On the Internet, you can be more census-based. You can measure every single video, which we do. So we now know about over half a billion videos and we've measured over three trillion video events.
Because you have this very deep census data of everything that's happened, you can use standard and interesting statistical processes to figure out exactly what's happening in that space, without having to extend a relatively small panel. You know what everyone is doing.
Gardner: And of course, this extends not only to programming or entertainment level of video, but also to the advertising videos that would be embedded or precede or follow from those. Right?
Meisl: Exactly. Advertising and video are interesting, because it's not just standard television-style advertising. In standard television advertising, there are 30-second spots that are translated into the Internet space as pre-roll, post-roll, mid-roll, or what have you. You're watching the content that you really want to watch, and then you get interrupted by these ads. This is something that we at Visible Measures didn't like very much.
We're promoting this idea of content marketing through video, and content marketing is a very well-established area. We're trying to encourage brands to use those kinds of techniques using the video medium.
That means that brands will tell more extensive stories in maybe three- to five-minute video segments—that might be episodic—and we then deliver that across thousands of publishers, measure the engagement, measure the brand-lift, and measure how well those kinds of video-storytelling features really help the brand to build up the trust that they want with their customers in order to get the premium pricing that that brand has over something much more generic.
Gardner: Of course, the key word there was "measures." In order to measure, you have to capture, store, and analyze. Tell us a little bit about the challenges that you faced in doing that at this scale with this level of requirements. It sounds as if even the real-time elements of being able to feed back that information to the ad servers is important, too.
Meisl: Right. The first part that you have to do is have a really comprehensive understanding of what's going on in the video space.
Visible Measures started with measuring all video that’s out there. Everywhere we can, we work with publishers to instrument their video players so that we get signals while people are watching videos on their site.
For the publishers that don't want to allow us to instrument their players, then we can use more traditional Google spidering techniques to capture information on the view count, comment count, and things like that. We do that on a regular basis, a few times a day or at least once a day, and then we can build up metrics on how the video is growing on those sites.
So we ended up building this massive database of video—and we would provide information, or rather insight, based on that data, to advertisers on how well their campaigns were performing.
Eventually, advertisers started to ask us to just deliver the campaign itself, instead of giving just the insight that they would then have to try to convince various other ad platforms to use in order to get a more effective campaign. So we started to shift a couple of years ago into actual campaign delivery.
Now, we have to do more of a real-time analysis because. as you mentioned, you want to, in real time, figure out the best ways to target the best sites to send that video to, and the best way to tune that campaign in order to get the best performance for the brand.
Gardner: And so faced with these requirements, I assume you did some proofs of concept (POCs). You looked around the marketplace for what’s available and you’ve come up with some infrastructure that is so far meeting your needs.
Meisl: Yes. We started with Hadoop, because we had to build this massive database of video, and we would then aggregate the information in Hadoop and pour that into MySQL.
We quickly got to the point where it would take us so long to load all that information into MySQL that we were just running out of hours in the day. It took us 11 hours to load MySQL. We couldn’t actually use the MySQL. It was a sharded MySQL cluster. We couldn’t actually use it while it was being loaded. So you’d have to have two banks of it.
You only have a 12-hour window. Otherwise, you’ve blown your day. That's when we started looking around for alternate solutions for storing this information and making it available to our customers. We elected to use HP Vertica—this was about four years ago—because that same 11-hour load took two hours in Vertica. And we're not going to run out of money buying hard drives, because they compress it. They have impressive compression.
Now, as we move more into the campaign delivery for the brands that we represent, we have to do our measurement in real-time. We use Storm, which is a real-time stream processing platform and that writes to Vertica as the events happen.
So we can ask questions of Vertica as they happen. That allows our ad service, for example, to have much more intelligence about what's going on with campaigns that are in-flight. It allows us to do much more sophisticated fraud detection. There are all kinds of possibilities that you can only do if you have access to the data as soon as it was generated.
Gardner: Clearly if a load takes 11 hours, you're well into the definition of big data. But I'm curious, for you, what constitutes big data? Where does big data begin from medium or non-big data?
Meisl: There are several dimensions to big data. Obviously, there's the size of it. We process what we receive, maybe half a billion events per day, and we might peak at near a million events a minute. There is quite a bit of lunchtime video viewing in America, but typically in the evening, there is a lot more.
The other aspect of big data is the nature of what's in that data, the unstructured nature, the complexity of it, the unexpectedness of the data. You don't know exactly what you're going to get ahead of time.
For information that’s coming from our instrumented players, we know what that’s going to be, because we wrote the code to make that. But we receive feeds from all kinds of social networks. We know about every video that's ever mentioned on Twitter, videos that are mentioned on Facebook, and other social arenas.
All of that's coming in via all kinds of different formats. It would be very expensive for us to have to fully understand those formats, build schemas for them, and structure it just right.
So we have an open-ended system that goes into Hadoop and can process that in an open-ended way. So to me, big data is really its volume plus the very open-ended, unknown payloads in that data.
Gardner: How do you know you're succeeding here? Clearly, going from 11 hours to two hours is one metric. Are there other metrics of success that you look to—they could be economic, performance, or concurrent query volumes?
Tell me what you define as a successful analytics platform.
Meisl: At the highest level, it's going to be about revenue and margin. But in order to achieve the revenue and margin goals that we have, obviously we need to have very efficient processes for doing the campaign delivery and the measurement that we do.
As a measurement company, we measure ourselves and watch how long it takes to generate the reports that we need, or for how responsive we are to our customers for any kind of ad-hoc queries that they want or special custom reports that they want.
We're continuously looking at how well we optimize delivery of campaigns and we're continuously improving that. We have corporate goals to improve our optimization quarter-over-quarter.
In order to do that, you have to keep coming up with new things to measure and new ways to interpret the data, so you can figure out exactly which video you want to deliver to the right person, at the right time, in the right context.
Gardner: Chris, we're here at the Big Data Conference for HP Vertica and its community. Looking down the road a bit, what sort of requirements do you think you are going to need later? Are there milestones or is there a road map that you would like to see Vertica and HP follow in order to make sure that you don't run out of runaway again sometime?
Meisl: Obviously, we want HP and Vertica to continue to scale up, so that it is still a cost-effective solution as the volume of data will inexorably rise. It's just going to get bigger and bigger and bigger. There's no going back there.
In order to be able to do the kind of processing that we need to do without having to spend a fortune on server farms, we want Vertica, in particular, to be very efficient at the kinds of queries that it needs to do and proficient at loading the data and of accommodating asking questions of it.
In addition to that, what's particularly interesting about Vertica is its analytic functions. It has a very interesting suite of analytic functions that extends beyond the normal standard SQL analytic functions based on time series and pattern matching. This is very important to us, because we do fraud detection, for example. So you want to do pattern matching on that. We do pacing for campaigns, so you want to do time series analysis for that.
We look forward to HP and Vertica really pushing forward on new analytic capabilities that can be applied to real-time data as it flows into the Vertica platform.