I'd like to explain an interesting phenomenon that we may see before in daily life and is related to topics in paper "Lessons from Giant-Scale Services".It mentions that in a large-scale application, read operations number is dominant to write operation, so companies may deploy database serving writing operations to different place from those for reading. Balance manager is used for picking the server for specific request when routing .
Sometimes when finish editing profiles on a site, we may see no changes happen immediately when viewing profile page. It may be confusing and leads to re-operations. And the reason is related to infrastructure design just mentioned. Suppose that the servers for profile-editing page( writing operation ) is on the west coast and for page-viewing is on the east side. When we finish editing and then want to view the results, we are still routed to east coast. But because the replication-delay, weird thing happens.
The solution to get around this by facebook is to set the time of write operation to cookies in browsers. When load balancer finds these time infos in cookie, it will compare these writing time with current time. If the slot is big enough, so replication is definitely finished and server on east coast will be used as usual. If not, we will be routed to west coast.
No comments:
Post a Comment