The following is a guest blog post from our friends at Cloudera. We recently did a case study with them describing how we use Apache Hadoop to process Big Data.
How Rapleaf Works Smarter with Cloudera
Because raising the visibility of Apache Hadoop use cases is so important, in this post we bring you a re-posted story about how and why Rapleaf, a marketing data company based in San Francisco, uses Cloudera Enterprise (CDH and Cloudera Manager).
Founded in 2006, Rapleaf’s mission is to make it incredibly easy for marketers to access the data they need so they can personalize content for their customers. Rapleaf helps clients “fill in the blanks” about their customers by taking contact lists and, in real time, providing supplemental data points, statistics and aggregate charts and graphs that are guaranteed to have greater than 90% accuracy. Rapleaf is powered by Cloudera.
Business Challenges Before Cloudera
Rapleaf established itself as a data driven business early on, collecting feeds from numerous sources to create a single, accurate view of each customer. By 2008, “we were processing data in a complex pipeline that involved an organic structure of many MySQL instances and queues,” explained Rapleaf’s co-founder and vice president of engineering, Jeremy Lizt. “As data volumes increased, that structure became unmanageable and expensive. It started getting difficult to perform the kinds of operations that we wanted to be able to do. It was no secret that this wasn’t going to scale.”
As part of its data synthesis, Rapleaf performs numerous processes and analytics to get the data it collects into a refined state. If they want to change an algorithm, that requires re-processing of all their data. Because of this, Rapleaf didn’t even consider migrating to a relational data warehouse. “If we had a stable set of data and had to do incremental updates, a traditional RDBMS might have been appropriate,” commented Lizt, who had read Google’s papers on MapReduce and learned about Hadoop as an open source implementation of the Google paradigm. “It was pretty straightforward: Hadoop was the clear opportunity and seemed to be the only option for us.”
The company made the move to Apache Hadoop and was running in production after 9-12 months. They managed the environment independently for a year before learning about Cloudera. “We were already invested in the promise of Hadoop and our feeling was that Cloudera would be good for the community and good for Hadoop,” said Lizt. Rapleaf decided to migrate to Cloudera’s open-source Hadoop distribution (CDH) due to its quality and stability, and soon signed on as one of the first Cloudera Enterprise customers.
Rapleaf processes the many feeds of data that it collects and synthesizes all of that data into a single, accurate view using Hadoop. Log messages are sent through Scribe and loaded into the Hadoop Distributed File System (HDFS). Log data is loaded into Hadoop every ten minutes, amounting to 1-2 TB each day. Other data sources load hourly or daily. Rapleaf has other jobs that run periodically on the logs to compute stats and make sure everything is running correctly.
After processing their data in CDH, Rapleaf puts it into a distributed hash table — an open source key value store built by Rapleaf called Hank — that can be queried with very predictable, fast response times. The company has 250 TB on 280 CDH nodes today, with capacity for up to 400 TB of raw, unreplicated data. MySQL is also still in use at Rapleaf, mostly to store Hadoop’s output.
Impact: Business-critical Reliability at Scale
Hadoop provides the foundation for Rapleaf’s business. “If something is critical to your infrastructure, it’s hard to articulate value. It’s like asking what your heart or your liver is worth,” commented Lizt. With that being said, the reliability and stability of Rapleaf’s Hadoop platform is imperative. This is the biggest benefit that Cloudera Enterprise offers to Rapleaf.
As an early Hadoop adopter, Rapleaf doesn’t rely heavily on the support provided by Cloudera but finds peace of mind in knowing it’s there. “Every time we’ve had an issue, we’ve had very fast support,” said software engineer Andre Rodriguez.
Lizt added, “Cloudera’s engineers are a talented bunch of people — they’re really intelligent, and we have confidence in their abilities. Whether it’s a glitch or just a question, it’s really helpful for us to be able to get a quick answer from someone at Cloudera who knows what they’re talking about.”
Impact: Operational Efficiency
Hadoop delivers a massively scalable data processing and storage platform that costs, on average, 10x less than traditional relational systems. But deploying Hadoop and keeping the cluster running at peak performance is no easy task. Rapleaf has found value in Cloudera’s ability to simplify the deployment, management and monitoring of Hadoop through support, services and the Cloudera Manager tool.
Rodriguez reported three main advantages offered by Cloudera Manager:
- Job-level statistics — “A lot of the things that I used to do manually before, I can just do through Cloudera Manager now,” said Rodriguez. “We can go back and see specifically what happened with jobs, get statistics from each job, and — perhaps most importantly — it keeps that information in the database. We actually use that data to compute other metrics so we can decide where to spend our engineering time.”
- Configuration management — Cloudera Manager provides explanations for what every configuration parameter means. “We could get that information before, but it wasn’t easy,” noted Rodriguez.
- Visibility into long-term trends — While Rapleaf had a very mature configuration management system before implementing Cloudera Manager, they had less visibility into long-term trends such as how things were performing over time. This is another key benefit that helps Rapleaf identify focus areas for their engineering efforts.
In summary, Rapleaf has built its business on Hadoop. Because it is a mission-critical component of Rapleaf’s infrastructure, the company relies on Cloudera Enterprise to ensure a stable, reliable and optimally performing Hadoop platform 24×7.