Tumblr was founded in 2007, which is currently the worlds largest light blogging site with more than 196 million blogs, 93 billion posts. Meanwhile, the site Blogging request has reached to 23,000 per second. Recently, the companys Web site reliability engineer Michael Schenck announced its architecture design on HighScalablity.
In Tumblr, blog is the largest part of the website traffic. In tumblelogs, the height cushioning performance is a very important feature. This is not easy given the high views / post ratio supported by Tumblr, Let’s check the supporting infrastructure of tumblelogs.
278 employees, six individuals responsible for Tumblrs perimeter (Perimeter-SRE), including a manager.
More than 2800 servers, less than 20 percent of which support for blog
23,000 blogging requests per second during peak period
6500 blog cache cleaning per second during peak period
More than 196 million blogs
More than 93 billion posts
Previous architecture segmentation based on mapping
In the earlier time, Tumblr was ran in a very small scale 1 active plus an alternate proxy server, as well as varnish nodes in the same configuration. Tumblr of this scale is very easy to manage, monitor and service, availability is very high, too. Then soon after, Tumblr had to crazily expanded to dealt with capacity limits brought by the users explosion.
Something we can learn:
1.Sometimes, the answer to the question probably arise earlier than your questions. When facing some scalability challenges, do not waste time on those problems you have overcome in other places.
2.Keep it simple. You inevitably may suffered from solving the scalability problem through complexity.
3.Understanding your hash function. Selecting a hash function is also equally important.