Architecting AOL’s Data Layer for Content Analytics – Ian Holsman, October 2010

In the presentation below at Hadoop World 2010, AOL’s Ian Holsman talks about AOL implementation of a new data layer for personalizing content and increasing clickthroughs on news pages, and how this led to AOL’s involvement with Big Data technologies.

Starting with a business goal – increase relevance of content to increase clickthroughs

Holsman starts off:

What the Data Layer project was about initially was trying to make sense of the data that we have coming in at AOL. Before AOL, I worked at CNET … and they [had done] a project where they had Top Stories (or most popular stories) at the top of various pages. And that was getting 1-4% clickthroughs …

And what they then decided to do was personalize that a bit. And they [increased the clickthrough percentage] from 4% to 20% clickthrough. … So the whole aim of this project was to try to implement this at AOL. So instead of going just for most popular stories, we tried to get more related stories.

So that’s the background on how [AOL’s involvement with Big Data] started. … This started in 2008, so we’ve been doing this for a while now. And it’s morphed from what we originally started into something much bigger.

Holsman continues:

So it started with a question, “Can we do better than a Top Stories link?” So like a said before, at CNET they did studies and they [increased their clickthrough rates on news stories] from 4% to 20%. And I thought we could do something like that [at AOL]. And so I put a proposal through.

What that morphed into in the business requirements was to increase recirculation of the pages – basically trying to get users to click through more. And also to improve the revenue per page. The COO of the time basically had a chat with Yahoo!, and he asked us why has Yahoo! been able to get much more higher value-adds on their pages? And one of the reasons for that is they know more about the user.

So the three major goals of this inititiave was (i) to get better [i.e. more relevant] ads on the page, (ii) get better reader engagement [with content on the page], and (iii) enable the user to click through [on a piece of content and discover other related content on AOL].

AOL at the time had 72 major properties. Most people probably only know about 3 or 4 of them. So ideally we’d be able to push you through to other places and you’ll start using more of the AOL network.

And what we translated the mission to was a Related Page module. Initially [the scope] was site-specific … so if you were on Shopping we didn’t push you to Real Estate. But the plan was to eventually make it network-wide.

Architecting the Data Analytics Platform v.1

So, how did AOL solution this? Again from Holsman:

Most of the stuff we did at the time was bleeding edge technology. … We got a guy to write some Javascript for us to start measuring things. That’s probably the first thing to get your toe wet, is you have to start measuring things, and start getting the data into the cluster itself.

We wrote a custom Apache module to do third-party cookies. So the problems we had were (i) getting the data, (ii) making sure we can identify the user across sites – so we created a custom module to create a cookie which is shared across multiple domains.

We wrote a custom load processing module to push the data every 15 minutes to a Hadoop cluster. And we wrote MapReduce jobs to get the data, crunch through it, and produce reports and MySQL databases with the aggregated data so other groups can use it.

Holsman adds that one of major aims AOL had when they began collecting data was around Privacy issues. He elaborates:

[We try to make sure that people’s personal data – i.e. people’s names, addresses, e-mail addresses] (a) isn’t collected, and (b) isn’t made available to anybody – internal and external.

So we tried to keep it anonymous. We basically decided to ditch the IP numbers completely. So if you look at our data collection, we use something called WOEIDs, which is geographic location. … So the IP number was basically never sent to disk anywhere.

Most of the stuff we do with data has a Privacy guy involved. And that’s probably important … when you’re dealing with large amounts of data, you have to think about the privacy. What happens if this gets out – if an internal user starts exposing it, or we have an Oops, and we have it [made public]? Especially with this level of data, and the amount of data you’re collecting.

The following diagram presents AOL’s initial architecture circa 2008 (sorry, I know it’s a bit small and hard to read):

Key elements of the architecture – split along East and West coasts – include:

• Beacons – collect interaction data. Beacons are provided by various analytics data gatherers/tools like Omniture and Comscore.
• Web Server Logs – AOL captures this data in Web Server logs, and sends the log data to the Hadoop cluster using the Hadoop protocol
• Hadoop Cluster – where data is processed using basically ETL-type transforms, that’s where all the jobs run
• Processed Analytics data – Data processed in the Hadoop cluster is sent to MySQL databases for real-time application access, as well as AOL’s Neteeza data warehouse for enterprise analytics reporting

This data flow – from Web Server to Hadoop to real-time MySQL databases, available for use by Web Servers – was happening every 15 minutes. AOL is currently redesigning the architecture to process this data in real-time (remember, this is 2008).

Holsman elaborates on the concept of Web Beacons:

If you ever look on a web page from a major web site, you’ll find that there’s various web collection “bugs” or beacon servers on their pages. So one of the ones we use is Omniture, and they give us page views. What this project was designed to do is grab this [beacon information] from Omniture and integrate it into our existing [infrastracture]. There’s also Comscore and various [advertising-related beacons] – it’s kind of scary how many beacons there are on most web pages.

Our initial goal was not to replace Omniture for page-view information, it was originally designed to collect “related site” information. We started to learn we might be able to replace Omniture with this infrastructure, but that was never our goal. There’s also advanced analytics things that Omniture does that we could start doing, but we’re not at that stage yet. That’s a big, probably multi-year, project to do that.

How AOL’s data team got started

Holsman goes on to talk about how they started. It basically started as a skunkswork project with a some spare machines lying around. Installed Hadoop on the servers. Installed a beacon on the Real Estate site, and started collecting data. It was important, says Holsman, that the data team didn’t wait for enterprise consensus from all the players in the organization. Rather, they started with a single channel – the Real Estate channel – and then began collecting data.

It took AOL 2-3 months to get the infrastructure installed and the data logs starting to come through. And then we started rolling it out to other sites across the enterprise. AOL now have all their 200+ sites feeding data into their data analytics platform.

Says Holsman:

At the time, Hadoop was pretty new to AOL. So we also used the Hadoop platform to build applications that the business people could see and derive value from. And then let the business drive [investment in the platform based on their needs].

When we looked at a web analytics platform for Bebo, it was cost-prohibitive to use Omniture. So we used our Hadoop cluster to track basic page views, unique views, and other basic information.

With the Bebo implementation, it became time to make this initiative a “real project” at AOL. At this point, AOL ran into a few problems. The most significant was that MapReduce was slow-to-write, and inflexible (note that this was before Pig was released, which greatly simplifies writing data transformations on top of the Hadoop platform). And Hadoop kept on hanging, which was a headache. The other issue the AOL team had was upgrading Hadoop – from 0.18 to 0.19, and then to 0.20. Holsman adds that most of this stuff has been fixed, and these issues are no longer a problem for AOL. But at the time they presented challenges.

People knowledgeable with working with Hadoop technologies was also a challenge. The AOL team didn’t have access to a consulting group at the time to assist them with their deployment. But this problem was also addressed through learning and training.

Yahoo! open-sources Pig

Then around 2008/2009, Yahoo! open-sourced Pig. Says Holsman:

Pig actually solved a lot of the people issues we had. It was much easier to use. Training was much simpler. And we could then basically push [development on the Hadoop platform] out to regular developers. Before Pig came out, we had a central team that [did all the application development on Hadoop]. And they were basically a bottleneck. We had 5 people that basically wrote MapReduce jobs all day … and there basically weren’t enough people to [properly service the business].

When Pig came along, we could then handoff the jobs to [other AOL developers], and they could write their own scripts and run it on our cluster. And what we then became – rather than a central processing house – was a data provider. We provided the data. We provided the machines [for the Hadoop cluster to run on]. We provided training for internal teams on how to use the stuff. And then we let them go wild.

The Result … Business Innovation

Holsman continues:

And “letting them go wild” was kind of risky, because it was like “Oh my godness, they’re going to hang the cluster …”

But what it actually did lead to was a lot of new innovations. I mean the channel developers are really smart guys in their own areas. They knew the business better than we did. We knew the data. They knew the business requirements.

So we basically opened it up, let them have access to their [data], and showed them how to use it. And they used it in ways that we never expected.

So the Hadoop Analytics platform basically became a source of innovation and product development at AOL. Here’s an example:

Again this is a bit hard to see. But the Auto Channel GUI designers are able to see a Heatmap of the page where users are clicking through, what links they are clicking on, and how good the page was. The Auto design team could then do A/B testing to see which pages produced better results. They could launch a new page, and within 15 minutes see where activity was happing on the page. And this was the #1 use case that prompted business-driven adoption of the Hadoop Analytics platform throughout the enterprise.

Analytics-driven Applications built on top of Hadoop Data

AOL has also built applications on top of Hadoop. For example, AOL developed a Shopping Recommendations site using Mahout machine learning and data mining library. Holsman elaborates:

At the time, we were looking at some [Shopping Recommendation] vendors. The Shopping site actually wanted to use an external vendor for this. We had two people at the time, and we wrote [a Shopping Recommendation Engine] internally. We used A/B testing to compare our results with 3rd-party results.

And we did better just using the algorithms that were available in Mahout, which we just downloaded. There’s no PhD’s working in the group. We understand Clustering to a certain degree. But we just downloaded the clustering algorithms, and just ran them. And they produced better results that what was available [from third-parties].

… We deployed the system on one site. Got it working. And now we can basically use the same algorithms on other sites.

AOL also built a User Recommendation capability (which had not been released at the time the talk was given) to recommend personalized news content to users leveraging the Hadoop data platform.

At the time of the talk, Holsman commented that AOL had the content side of the platform working. And that AOL was currently working to integrate Advertising and Lifestream platforms into an overall Analytics/Targeting platform.

Moving Forward …

Here are the current goals for the Data Analytics team at AOL:

1. Get more information about our customers
2. Build metrics into our platform
3. Build intelligence on the page – for example: Collaborative Filtering, Product Recommendations, Top-K Type Lists
4. Make the analytics platform closer to real-time

AOL’s Data Analytics Infrastructure today (circa 2010)

Here’s a diagram of AOL’s data layer infrastructure today:

Elements that are included in this infrastructure that are not seen in the 2008 version include:

• Publishing Platforms
• Advertising Web Servers – in addition to web servers that deliver content
• Relegince – AOL’s in-house semantic content platform
• Pig – For writing data flow/transformation scripts (not shown in this diagram)
• 2 Cassandra databases – 1 for storing and servering real-time user information, and 1 for storing some type of clustering information
• Redis Key-value data store – not sure what it’s place in the architecture is

In Summary

A very insightful talk! A quite fascinating glimpse into how AOL is architecting a near real-time semantic content and ad-serving platform, as well as a data analytics platform that powers the real-time semantic content platform.

glenn

Facebook’s Architectural Stack – designing for Big Data

This is my fourth of a series of posts exploring the topic of Big Data. The previous posts in this series are:

• Designing Large-Scale Retrieval Systems at Google – Jeff Dean from 2009
• Introduction to Hadoop – understanding Big Data
• Keen Insight on Big Data from Cloudera CEO Mike Olson

This post provides two videos in which Facebook’s David Recordon discusses Facebook’s architectural stack as a platform that must scale to massive amounts of data and traffic. The first video is a short video where Recordon discusses Facebook use of the LAMP stack at OSCON 2010:

On Database Technology and NoSQL Databases at Facebook

In the first video, Recordon first addresses how Facebook implements database technology generally, and the topic of NoSQL databases. Says Recordon:

The primary way that we store data – all of our user data that you’re going and accessing when we’re working on the sight, with the exception of some services like newsfeed – is actually stored in MySQL.

So we run thousands of nodes of a MySQL cluster – but we largely don’t care that MySQL is a relational database. We generally don’t use it for joins. We’re not going and running complex queries that are pulling multiple tables together inside a database using views or anything like that.

But the fundamental idea of a relational database from the ’70s hasn’t gone away. You still need those different components.

Recordon says that there are really three different layers Facebook thinks about when working with data, illustrated in the following visual:

Continues Recordon:

You have the database, which is your primary data store. We use MySQL because it’s extremely reliable. [Then you have] Memcache and our web servers.

So we’re going and getting the data from our database. We’re actually using our web server to combine the data and do joins. And this is some of where HipHop becomes so important, because our web server code is fairly CPU-intensive because we’re going and doing all these different sorts of things with data.

And then we use Memcache as our distributed secondary index.

These are all the components that you would traditionally use a relational database for:

Recordon continues:

[These are the same layers that were] talked about 30-40 years ago in terms of database technology, but they’re just happening in different places.

And so whether you’re going and using MySQL, or whether you’re using a NoSQL database, you’re not getting away from the fact that you have to go and combine data together, that you’re needing to have a way to look it up quickly, or any of those things that you would traditionally use a database for.

On the topic of NoSQL databases, Recordon says:

And then when you dig into the NoSQL technology stack, there are a number of different families of NoSQL databases which you can go and use. You have document stores, you have column family stores, you have graph databases, you have key-value pair databases.

And so the first question that you really have is what problem am I trying to solve, and what family of SQL database do I want to go and use.

And then even when you dig into one of these categories – if we just go and look at Cassandra and HBase – there are a number of differences inside of this one category of database. Cassandra and HBase make a number of different tradeoffs from a consistency perspective, from a relationship perspective. And so overall you really go and think about what problem am I trying to solve; how can I pick the best database to do that, and use it.

While we store the majority of our user data inside of SQL, we have about 150 terabytes of data inside Cassandra, which we use for Inbox search on the site. And over 36 petabytes of uncompressed data in Hadoop overall.

On the topic of Big Data

Recordon:

So that leads me into Big Data. We run a Hadoop cluster with a little over 2,200 servers, about 23,000 CPU cores inside of it. And we’ve seen the amount of data which we go and store and process growing rapidly – it’s increased about 70 times over the past 2 years. And by the end of the year, we expect to be storing over 50 petabytes of uncompressed information – which is more than all the works of mankind combined together.

And I think this is really both the combination of the increase in terms of user activity on Facebook … But also just in terms of how important data analysis has become to running large, successful websites.

The diagram below shows Facebook’s Big Data infrastructure:

Says Recordon:

So this is the infrastructure which we use. I’ll take a minute to walk through it.

With all our web servers we use an open source technology we created called Scribe to go and take the data from tens of thousands of web servers, and funnel them into HDFS and into our Hadoop warehouses. The problem that we originally ran into was too many web servers going and trying to send data to one place. And so Scribe really tries to break it out into a series of funnels collecting this data over time.

This data is pushed into our Platinum Hadoop Cluster about every 5-to-15 minutes. And then we’re also going and pulling in data from our MySQL clusters on about a daily basis. Our Platinum Hadoop Cluster is really what is vital to the business. It is the cluster where if it goes down, it directly affects the business. It’s highly maintained, it’s highly monitored. Every query that’s being run across it, a lot of thought has gone into it.

We also then go and replicate this data to a second cluster which we call the Silver Cluster – which is where people can go and run ad-hoc queries. We have about 300 to 400 people which are going running Hadoop and Hive jobs every single month, many of them outside of engineering. We’ve tried to make this sort of data analysis to help people throughout the company make better product decisions really accessible.

And so that’s one of the other technologies which we use, Apache Hive, which gives you an SQL interface on top of Hadoop to go and do data analysis. And all of these components are open source.

So when Facebook thinks about how there stack has evolved over the past few years, it looks something like this:

Where the major new component is the Hadoop technology stack and its related components to manage massive amounts of data, and do data analysis over top of that data.

A deeper look at Scaling challenges at Facebook

The second video is a presentation delivered by David Recordon and Scott MacVicar – both Facebook software engineers – at FOSDEM in February 2010 provides a deeper look into Facebook’s use of open source technology to provide a massively scalable infrastructure:

The question that I am interested in, and isn’t answered in these videos, is how Facebook implements its Open Graph data model in its infrastructure. That would be very interesting to learn. For more specifically about Facebook’s Open Graph technology, please see Facebook’s Open Graph and the Semantic Web – from Facebook F8.

Very interesting stuff.

glenn

Keen Insight on Big Data from Cloudera CEO Mike Olson

Interesting insights on the origins of, and trends in, Big Data from Cloudera CEO Mike Olson:

For more on Big Data, see my previous post: Introduction to Hadoop – Understanding Big Data.

glenn

Introduction to Hadoop – understanding Big Data

This is my second post on my journey of understanding Big Data. My first post looked at the design of large-scale retrieval systems at Google. This post looks at Hadoop – a framework for processing massive data sets across multiple nodes, insipired by Google’s MapReduce and GFS architectures.

There are two nice video introductions to Hadoop – both sponsored by O’Reilly Media and both featuring Tom White, author of Hadoop: The Definitive Guide. The first webcast is from July 2009 titled An Introduction to Hadoop:

The second webcast is from September 2010 titled The State of Hadoop:

For a wonderful overview of the state of Big Data, please see Making Sense of Big Data, a PWC Technology forecast from 2010.

More to come on Hadoop and Big Data in future posts.

glenn

Designing Large-Scale Retrieval Systems at Google – Jeff Dean from 2009

Starting to tackle the topic of Big Data. And the place I’m starting is two videos from a few years back, where Google’s Jeff Dean discusses the architecture of large-scale retrieval systems at Google. Here’s the first video, from March 2009:

Vodpod videos no longer available.

And here’s the second video from August 2008:

For more on the topic of Big Data, please see the following articles:

• Jeff Jonas: Big Data is the Next Big Thing – TechCrunch TV, October 2010
• Making sense of big data – PWC Technology forecast, Issue 3 2010
• A conversation with IBM’s Mr. Big Data – InfoWorld, February 2011

glenn