Sphinx - text search The Pirate Bay way
Like MySQL. But it can scale
In 2001, Andrew Aksyonoff had an itch. He ran a website that indexed song lyrics and allowed users to search them. At the time, none of the open source options for text search worked. Lucene, the standard response at an engineering team meeting when the problem of search comes up, was still a baby. Indexing took too long, and it was a bear to install on anything other than a Lucene developer's machine.
Mnogosearch - the search engine you probably still haven't heard of - was too bloated and took a full twenty-four hours to index Aksyonoff's meager 150 megabytes of data. On top of all this, the standard algorithms for text search were about as useful as a venereal disease when it came to searching song lyrics. Andrew had a different idea of how text search should work - and the spare time to try it out. He did, and today, his software powers the search engines behind Craigslist, The Pirate Bay, and many more.
It's called Sphinx, and it's on track to become the open source world's canonical answer to the question of text search. MySQL and Solr, the two popular solutions, are showing their age. MySQL introduced full-text search in late 2000 as a way to more intelligently search blobs of text stored in databases. You can work a full-text clause into a query, and MySQL will rank the result rows by how relevant it thinks they are to the query.
MySQL uses textbook search algorithms and doesn't allow for a lot of relevance tuning. It's like a drawing from a five year old: The heart is in the right place, but everybody knows that kids suck at drawing. Implementation details aside, MySQL still suffers from scalability problems. Having ignored the trend of chip manufacturers to build multiple cores into CPUs, hoping that this unpleasant trend that required them to actually think about multi-threading would just blow over sooner or later, MySQL's ability to handle parallelism is, well, see the five year old's drawing.
Internet-famous MySQL wonk Jeremy Zawodny, who had the foresight to jump from the ship's bow as Yahoo started to take on water, replaced MySQL fulltext search at Craigslist with Sphinx. Craigslist used 25 machines to handle roughly 50 million queries per day on MySQL. Under that kind of load, Zawodny found that MySQL wasn't using much CPU or doing much disk I/O, which means it's spending all of its time waiting on thread locks. Oops.
Maybe we should have paid attention to parallelism after all. The Sphinx implementation took those 25 machines down to 10, with plenty of room to grow. While Sphinx didn't handle the traffic out of the box at the time, Zawodny was able to patch it to handle Craigslist's specific need - and fix a few bugs along the way.
Vowels are a scare resource
Before Sphinx, the other option for text search was Apache Solr. Solr, whose name-giver understands that vowels are a scarce resource and must be used sparingly, is a server that sits on top of Lucene. Solr is popular with the enterprise crowd, who love its Java. Being a Java program, Solr includes no shortage of technology whose acronyms contain the letters J and X.
This tickles the enterprise pink, because these sorts of developers love nothing more than hanging out around a whiteboard drawing boxes and arrows and, from time to time, writing XML to make it look like they're doing real work. Solr thrives in this environment, being an Apache Foundation project, the Apache Foundation, of course, widely known as a cruel experiment to see what happens when bureaucrats do open source.
Being backed by Lucene, Solr is bound by Lucene. It is easier to tune Solr for relevancy than it is to tune MySQL, but it still doesn't provide the search quality to justify its popularity. That, and Solr has scalability problems. Lucene was originally designed as an embeddable library, so scalability wasn't a top priority. As a network layer on top of Lucene, Solr must accommodate for this. "In my internal benchmarks, Solr routinely performed 2x-4x slower [than Sphinx], both for indexing and searching," Aksyonoff tells The Reg.
Sphinx takes a different approach to search. For example, using MySQL or Solr, each query will be trimmed of its stopwords. This is a method out of the information retrieval books, where you remove low-signal words (like "the" or "and") that can degrade search quality. This was one of Aksyonoff's original problems searching song lyrics, a query like "I feel you" would yield Cuil-like results. By default, Sphinx does not remove stopwords (but it can be configured to) and ranks results based on phrase matching.
Having its roots in database search, Sphinx speaks MySQL, which means you can connect to its search server with a MySQL client and issue queries that look very similar to SQL. It will easily index data out of a MySQL or Postgres databases, but also supports an XML format for arbitrary data. There are also clients for PHP, Python, Ruby, and any other language that's worth using. Scalability is no issue, both in searching and indexing: Sphinx can index 10 megabytes of data per second and can search up to 100 gigabytes of text on a single processor. It also supports multi-machine distributed searching, as in the case of Craigslist.
In the vicious meritocracy of open source software, the only yardstick of success is the distribution of your code. In this regard, Sphinx still has some way to go, but there's still profit to be made. High-volume installations at Craigslist and The Pirate Bay are a validation for Aksyonoff, who now makes his living on Sphinx.
"We aren't exactly bathing in money, but on the other hand, we're paying the salaries from our revenues instead of burning venture capital and have always been," he says. His company, Sphinx Technologies, provides consulting and support services, as well as custom development. Judging by how well Sphinx is performing in real-world installations, for Aksyonoff, the General Public Business Model might actually pay off.
You can download Sphinx here. ®
Ted Dziuba is a co-founder at Milo.com You can read his regular Reg column, Fail and You, every other Monday.