September 16, 2013
At Gilt Groupe, each day at 12pm is peak site traffic time and their main user store receives 100,000 requests per minute. To handle these traffic spikes, they need a durable database that will stay up no matter what. They decided to look into using Riak for this main user store but, before they could bring it into production, they needed to performance test it.
With the help of two Riak engineers, Seth Thomas and Steve Vinoski, they were able to create a Riak test system in a week. They stress tested Riak and Riak passed with high marks. Riak also continued to perform well during a number of staged failure scenarios.
To ensure that Riak was actually as solid as the testing seemed to prove, they even simulated some extreme read and write loads. Zeusbench executed an extreme number of HTTP requests against the test service and reported metrics on performance. They also wrote a program that executed 1,500 insertions into Riak per second. Even with both these running, read/write performance in Riak never suffered.
Gilt Groupe was also able to kill multiple nodes in Riak while maintaining the production load. Riak continued to shine.
For more information about how Gilt Groupe tested Riak, check out their blog. Additionally, if you’re in New York tomorrow (September 17th), register for the Riak Meetup at Gilt Groupe’s office. Jim Englert (Gilt Groupe Lead Software Engineer) will be speaking about his experiences stress testing Riak and Riak’s Tom Santero will talk about what’s in store for the future of Riak.
