The Umbrella Security Labs research team has been sharing frequently about how they’re leveraging Big Data to predict unknown threats. And, the OpenDNS and Umbrella product teams have been working to improve the quality and speed of reports in our user Dashboards. The foundation of each of those discussions is how our infrastructure team handles the data itself. This post is an exploration of how OpenDNS handles the massive amounts of data we process daily, without downtime or performance impact, and what it means for the reporting in the Umbrella Dashboard.
On an average weekday the DNS resolvers we run at OpenDNS process more than 50 billion queries originating from over 50 million individual IP addresses. These queries are directed at our anycast IPs (220.127.116.11 and 18.104.22.168), which get routed to one of our 20 data centers, and from there to one of our 80 individual resolver hosts. All combined, this results in a huge amount of data. That data must somehow be processed and aggregated to produce the reports available on our users’ Dashboards, and sampled by Umbrella Security Labs who analyze the data to detect and predict security threats.
Each query that we log on our resolvers includes the domain being queried, the query’s originating IP, appropriate customer IDs associated with it, and how the query was handled (i.e. whether it was handled normally or blocked due to malware or phishing). These log entries average around 115 bytes each. Our system produces between 50MB/s and 90MB/s, for a total of over 5TB of raw data produced every day.
When I started working at OpenDNS in July of 2010 we were only receiving around half the queries we do now, but the analysis system in use then had been built years earlier. It was barely keeping up. It clearly wasn’t going to scale much further. Like many other companies that have delved into “Big Data” analysis, we built a Hadoop-based infrastructure to replace the previous installation, and it went live in early 2011.
Our production Hadoop cluster is composed of roughly 30 (and rapidly growing!) heavy-duty Ubuntu servers. Of these, one is the active Namenode, the cluster’s coordinator and Hadoop’s only single point of failure. To alleviate this risk, we synchronously replicate the Hadoop metadata on this machine to a spare Namenode via DRDB. That allows us to rapidly failover to the spare machine if necessary, and provide a means to update the Namenode without significant downtime. The remaining worker nodes are basically interchangeable and new ones can easily be added as demand increases.
On our resolvers the query logging is configured to roll over every 4MB, which results in a new log file every few seconds on each resolver. Sitting between our Hadoop cluster and the resolvers is a set of “loader” machines which pull in these completed log files, combine them into larger chunks, compress them using LZO compression, and finally push them to the Hadoop Distributed File System (HDFS). HDFS replicates data across multiple nodes for redundancy (in our case there are three copies of all data in HDFS), and as the compressed resolver logs for a day currently average 1.1TB, this allows our system’s ~300TB of storage to hold 3 months of raw data.
Once all the logs for a given hour have been received, our system launches a variety of Map/Reduce jobs to crunch that data and generate aggregate data for our customer-facing reports. The output of these jobs are loaded into HBASE (a distributed No-SQL database built on top of HDFS) for future access. When a customer visits the reporting page of the OpenDNS Dashboard, the Dashboard sends a query to one of several Appservers which query HBASE for the time frame requested, sum up the data, and return it to the Dashboard for presentation.
In order to provide cluster-level failure handling, and be able to upgrade the production cluster while continuing to return reports to the Dashboard, we have an additional cluster which does nothing but provide a live backup of our HBASE data. If the production cluster has failures, or needs to be taken offline, our Appservers can quickly switch to querying the backup HBASE. As an additional level of paranoia, we also take a nightly snapshot of this backup HBASE which we copy to a machine outside of the Hadoop cluster.
And that’s the mile-high overview of the infrastructure we use to analyze the massive amount of data our system produces. Our infrastructure is also rapidly evolving as we expand our security research team, develop new reports, and work towards providing real-time analysis and reporting capabilities. If you enjoyed reading about how the OpenDNS Global Network’s infrastructure powers Umbrella, I encourage you to read more in our recent whitepaper. The paper takes a look at how Umbrella Security Labs is harnessing the massive data sets mentioned above to predict future threat origins.