Processing the Deep Backlist at the New York Times

At the O’Reilly Open Source Convention (OSCON), Derek Gottfrid of the New York Times led a fascinating session on how the Times was able to utilize Amazon’s cloud computing services to quickly and cheaply get their huge historical archive online and freely viewable to the public.

How big is the archive? Eleven million individual articles from 1851 to 1980, or 4 terabytes of data (over 4,000 gigabytes). The Times got it ready for distribution in 24 hours, for a total cost of $240 in computing fees and $650 in storage fees.

As part of their original TimesSelect subscription service, the paper had scanned their entire print archive. Each full-page scan was cut into individual articles. Typical of newspaper format, the articles often spanned column or page boundaries, which meant that many articles were composed of several scans. In the original subscription-based program, whenever a reader requested one of these historical articles, the Times computer would need to stitch together all of the scans for a particular article before presenting it.

This on-demand process used significant computing resources, but because TimesSelect was subscription-based there was never much traffic. Once this archive was open to the public it was expected to generate greater usage, and the safest approach in those cases is to serve pre-generated versions of all 11 million articles. Using traditional software development practices — with a single computer churning through one article at a time — the processing could potentially take weeks and tie up Times servers that were needed for other tasks.

Gottfrid turned to Amazon Web Services (AWS) and its two main products:

Amazon Elastic Compute Cloud (EC2) is a form of “virtualization” where one very large computer is divided up into many virtual computers that can be individually leased out for use. Traditional hosting costs money whether the server is working or idle; in EC2 you pay only as long as the virtual computer is running. When it’s no longer needed, it’s shut down. This makes the service ideal for one-off processing jobs.

In addition, Amazon doesn’t care whether you use one EC2 “instance” 100 times, or 100 instances all at once — the cost is the same. The difference is when you can usefully divide a job into 100 concurrent tasks, because then it takes 1/100th the total time.

Amazon’s other major AWS offering is the Simple Storage Service (S3), for large-scale file hosting. Like EC2, it is a leased model — you pay only for the space that you use in a given time period.

Gottfrid leveraged these technologies in combination with a relatively new software library called Hadoop. Hadoop is written in the Java language and is based on work done at Google. It allows programmers to very easily write programs that can be run simultaneously on multiple computers.

Combining Hadoop concurrency with EC2 and S3, the Times was able to run a job that might have taken weeks of processing time and complete it in 24 hours, using 100 EC2 instances. They were pleased enough with S3 it became their permanent hosting platform for the scans. Hosting with Amazon or other cloud computing services is usually cheaper and has much better bandwidth than the average provider, although downtime can and does occur.

At last year’s OSCON, the Times announced the formation of its developer blog, Open. You can read more about the original AWS project as well as TimesMachine, a project that became economically feasible due to the low cost of AWS.

tags: , , , ,