Ole Markus has a new post to his blog looking at a technique for working with memcached and fetching data out of the store using a binary semaphore for better performance.
A typical document takes but a few hundred milliseconds to generate when a single request for the document enters the backend. The problem is that this is a highload website. In its current form, the backend serves hundreds of pages per second. This pretty much guarantees that the backend will concurrently receive cache miss on multiple languages and at the same time also receive cache miss on the pre-translated document.
Given that he wants the translated version to be the one that's always shared, a problem can come up when the cache request is "missed" and the document starts generating from multiple places. His fix for the situation is that only the first miss generates and all others see a lock on it and wait for it to be removed before successfully fetching the result. He provides code in a "LockedMemcached" class to help make it all more useful.