Unfortunately I just checked and it seems nearly half the traffic recorded by the mirror redirection service (we log the requests including the size of the file so we can calculate how much traffic we are sending to mirrors) over the last couple of days was non-redirectable requests to the dists directory we average about 66Mbps of such traffic from the mirror redirection system. I'm sure it wasn't anywhere near such a high proportion when I checked previously (it may be an anomoly, caused by a recent blog post on the rpf site).> The main thing with mirrorbrain is you can't safely use it on the dists > directory because of it's lack of mechanisms for correctly handling files > that change in place. We handle this simply by serving the dists directory > off the main server without any redirection obviously this comes at a price > in bandwidth from the main server.. That seems like a pretty good tradeoff though, as the bulk of the data transfer will be binary packages.
Note that we do not use pdiffs because of the added complexity and because in our experiance they are painfully slow on arm systems. We also don't split main and security suites (again for simplicity, importing the same package into multiple suites creates all sorts of build-management complications). If we used pdiffs and/or split main/security suites (with the main stable suites rarely updated) then I would expect the figures to be somewhat different.