How Digg.com uses the LAMP stack to scale upward
- — 26 April, 2007 08:31
Digg.com credits two particular features of its LAMP (Linux, Apache, MySQL and PHP) server cluster for helping the news aggregation site maintain speedy performance in the face of high growth.
The site, which lets its users vote on, or "digg," their favorite news stories hosted on other sites, recently passed the 1.2 million-user mark according to Elliot White III, an engineer at San Francisco-based Digg. He spoke at MySQL's annual conference in Santa Clara, California on Tuesday.
Today, Digg.com boasts 100 servers scattered in multiple data centers that host a total of 30GB of data, but the site started off in late 2004 as a single Linux server running Apache 1.3, PHP 4, and MySQL 4.0 using the default MyISAM storage engine, White said.
As more users dug Digg, the site moved to an architecture that uses a load balancer in the front that sends queries to PHP servers, MySQL slave servers that feed the PHP servers, and a MySQL master server that feeds data to the slaves.
That's a fairly standard setup. But to get away from "sending raw queries against the database," White said Digg.com uses a software called Memcached. First developed for use by the Livejournal site, Memcached is tailored for dynamic sites like Digg.com, which serve Web pages with content that is constantly changing and is personalized according to user preferences, White said.
Memcached stores chunks of data that can be pulled and used to dynamically create a Web page. Conventional caching systems, which store whole Web pages, would be too slow and inefficient for a site like Digg.
The other atypical feature of Digg's setup is its use of what Tim Ellis, another Digg engineer, calls "sharding."
A term apparently coined by Google engineers, sharding involves breaking a database into smaller parts in order to isolate heavy loads for better performance.
"If 90 percent of your data is within a certain range, and you can get that part working really fast, then you can help customers," Ellis said. "Then it's OK if the remaining 10 percent is slower."
A database can be sharded by table, date or range. It is similar to partitioning, says Ellis, but with several key differences. Sharding usually involves divvying up data onto different physical machines. Partitioning, in contrast, typically occurs on the same piece of hardware. And while MySQL does not natively allow sharding, it does support partitioned tables, federated tables and clusters.
While sharding has helped Digg.com achieve much faster performance overall, breaking a database into several smaller ones increases complexity, Ellis said. That can mean more work for developers and database administrators, because of the inability to use common SQL commands such as joining tables. "Developers don't like this crazy stuff. That can create pushback," he said.
Digg's current architecture includes about 20 database servers, 30 Web servers, and a few search servers running Lucene; the balance operate as backup servers. All but one of the database servers run some version of MySQL 5. The transaction-heavy servers as well as the backup units use the InnoDB database engine, while the OLAP ones use MyISAM.
Ellis acknowledges that Digg.com "is really lucky" in that 98 percent of the time the database is accessed, it is being read, as opposed to experiencing more intensive data writes.
"Most people come to Digg's front page, read it and leave, which is kind of nice," said Ellis, drawing a knowing laugh from the audience of mostly PHP developers and DBAs.
Ellis also noted that although many users have complained that upgrading to MySQL 5 from 4.1 caused performance to drop, that was not true in Digg.com's case.
Maintaining Digg.com's high performance as the site grows more and more popular presents challenges to Digg engineers. For one thing, the company is unable to keep scaling by buying more physical memory. "We can't afford that anymore," White said.
Preventing Digg's enthusiastic developers from adding powerful but CPU-intensive features is "a political thing I constantly have to deal with as a DBA," said White.
Also, Digg was having a problem with its storage misreporting the status of data synchronizations. "Our hardware wanted to be fast," White said. "It was telling us things were synced to disk when it was not."
Finally, there is the mundane challenge of minimizing "schema cruft," or redundant tables of data which, if read, can slow down performance, said White.
"Everyone has to do this," he said.