On Fri, Oct 26, 2007 at 09:25:19AM +0200, Klaus Kaempf wrote:
* Dr. Peter Poeml <poeml@suse.de> [Oct 25. 2007 17:45]:
On Thu, Oct 25, 2007 at 05:39:12PM +0200, Klaus Kaempf wrote:
Am I right to rephrase "it is redirected to the same mirror again on the next request" as "the redirector is introducing the stickyness" ?
Yes, and it is optional behaviour, which can be switched off in the Apache module. --Which is what I meant with "we have been running with and without stickiness in the past".
So at least the "retry button doesn't switch mirror" problem mentioned by Benji could be solved in the Redirector until the client tools catch up.
That doesn't solve anything. It only alleviates the problem, on the cost that it hits much more people (like nearly everybody). Not a good trade. And as a side effect, it makes it nearly impossible to track down problems with mirrors. Thus, it increases the damage. No, it is important to find out about broken mirrors as soon as possible. I have been through it: I tracked down a "broken mirror" during the time we ran *without* stickiness. A mirror had a funny firewall which partly swallowed some server responses. It took *weeks* until I had tracked it down. At first, there were only sporadic reports that "something didn't work", and you had no idea where to start looking. Fortunately, it happened to myself sometimes, because it was a German mirror. Note that a broken mirror would be chosen irregardless of stickiness, randomly. And the fewer mirrors a country has, the more often that happens. Stickiness helps to track down problems and to fix them. ... Of course, not all users even *think* of notifying us in case they encounter problems with a mirror. Instead, they might just keep asking around "how they can force another one". Still, if we get a report, or somehow find out about it by other means, we can do something about it pretty quickly. What we also need, is more active monitoring of mirrors, which includes issuing real requests to files. Just a few days ago, I have seen for the second time a mirror where 80% of requests where handled normally, while 20% were truncated. That could be easily detected with a script checking mirrors for correct and reliable replies by means of test requests. It would help if clients show how they follow through redirects, because that could make it apparent to users which mirror is causing trouble. Right now, it requires debugging capabilities to find out which server the request went to. Peter -- "WARNING: This bug is visible to non-employees. Please be respectful!" SUSE LINUX Products GmbH Research & Development