On Tue, October 31, 2006 12:43 pm, Roy T. Fielding wrote: > You say that the strategy will degrade performance, but you > don't indicate why. If there is some great performance penalty > caused by changing the interface to send_body, I'd like to know > what it is.
Changing the interface to a pure memory buffer interface removes the ability to "set aside" a brigade. The cache at this point therefore offers no choice as to write strategy to cache providers - it always writes disk network disk network, and if the network is slow, then so is the disk write, and so is the read from backend. A significant optimisation comes about when you try and read as much of the backend as you can, making the expensive CGI/proxy/mod_perl/whatever source "go away" as soon as possible. What do you do with all the data you just read if you can't immediately send it to the network? Put it in the cache of course. In addition, the ability to serve partially cached responses becomes impractical, which is one of the key drivers for large_disk_cache. Regards, Graham --
