We have weird traffic pattern for one product which is pretty much random request for very big (several TB) content base. Running traditional cache in this situation is impossible because Linux/UNIX file system (pretty much any kind of) hits the ceiling when d-entry cache fills memory. Directory lookups becomes a real issue when number of files reaches 6-7 digit numbers. Thats where custom file storage comes to help. After trying COSS with little success (it was not quite stable in 2.6) I tried varnish but inability of varnish to keep persistent cache (which is fixed right now I believe) made it unattractive. So I returned to COSS after 2.7 came out and it appeared to be working great and in fact it was so much greater than native storage that I forgot my experiments with varnish and run squid right now with 4-digit number of requests per minute per server with almost no server load and manageable disk IO. I will probably revisit varnish someday when persistent cache gets into stable state - but beware - you need serious kernel tuning before you turn varnish into production, by default it tends to dump dirty pages in very big chunks during which it does not serve requests....