Eschewing popular choices such as XML, CSV and JSON, Twitter has opted to format the back-end storage of its user and systems data with a relatively unknown format pioneered by Google, called Protocol Buffers.
With the company storing 12TB of this data each day for later use, the decision of which format to use was a crucial one.
"Getting your data formats right is everything," said Twitter analytics lead Kevin Weil, during a talk at the HadoopWorld conference in New York earlier this month.
The company is planning for the time when it will have to house "a trillion Tweets", Weil said, and it wants tools in place to analyse this information. The combination of Protocol Buffers, along with Hadoop and other associated technologies, should streamline this job.
When stored, each short message, or 'tweet', consists of 17 fields, six of which have at least one subfield. And the company will probably add more fields to these schema in the years to come.
In addition to the tweets the company's users supply, Twitter keeps internal log data on more than 80 different types of operations that occur within its systems. Much of this log data is aggregated by Facebook's open-source technology Scribe.
The choice of a format to store all this data was a difficult one. One obvious choice is XML, but that protocol is "very wordy", Weil claims, referring to how the name of the tag accompanies each data element.
Under XML, "one petabyte for a trillion Tweets might become 10 petabytes for a trillion Tweets," he said.
At the other end of the spectrum is CSV (Comma Separated Values). As the name suggests, CSV separates each data element only with a comma. While simple, it is not good for nesting data elements in subfields, Weil suggests. Also, if the schema is changed, the resulting programming it would take to accommodate data in the old schema would be considerable.
Protocol Buffers, used widely within Google, is an extensible protocol for serialising data, one Google claims is simpler than XML. And it can automate the process of recreating the data structures within applications.
"You define how you want your data to be structured once, then you can use special generated source code to easily write and read your structured data to and from a variety of data streams and using a variety of languages," a Google tutorial on Protocol Buffers states. "You can even update your data structure without breaking deployed programs that are compiled against the 'old' format."
For Twitter, this automation would allow the company to spin up new features more quickly.
"Protocol Buffers will generate code in a number of different languages, so you don't have to write code beyond IDL," or Interface Description Language, Weil said. It also ensures that should the schema be changed, the older information will remain accessible.
While primary copies of user Tweets are kept in MySQL and Cassandra databases, the company is also building a second data repository, running on Hadoop, that can be used for analytics and applications.
The information in this system can be queried using Java MapReduce or Pig, which is Hadoop's own SQL-like query language. Already one feature, Twitter's name search, runs on this system and more are expected to be built.