On 22 September a farm of automated browsers requested our public pages from thousands of addresses, and the site slowed to a crawl at about three pages a second. The cause was ours: every translated word on a page cost a separate round trip to a cache server, more than two hundred for one page. Those lookups are now answered from memory. Measured in production afterwards, the median page renders in 42 ms instead of 249, and the slowest in twenty in 72 ms instead of nearly two seconds. An overflow now gets a quick "try again in a moment" instead of slowing every page down.