When experimenting with a MirrorBrain setup that uses a dummy file tree, I ran into the
situation that the file tree wasn't complete, and I got 404s (file not found) in the
client. The same would happen if the tree is not up to date, and some new files are not
When trying to keep the system running under adverse circumstances, it doesn't make sense
to error out in such a case, and it would probably make sense to redirect such requests to
one of the fallback servers. (Referring to the fallback servers that can be configured
since recently, r7880.) Or maybe a different set of servers, don't know.
An obvious disadvantage is that those fallback servers end up getting _all_ requests that
requests that lead to a 404. Those mirror servers must be assumed to be fairly complete
for the whole thing to make sense.
On the plus side, this way the redirector could keep running even when it looses its file
tree (disk crash).
Not to forget, this feature (and similar ones) could be made configurable, so the
behaviour could be switched on only in emergency, thereby minimizing negative
consequences. Or, touching a file in the filesystem could signal to Apache that it needs
into "degraded mode".
As a slight variant of this, Apache could still do database lookups, even if the file tree
is gone. That would preserve the ability to redirect to all mirrors that have a requested
file, and only those that have it (and not blindly).
The feature would need to hook in earlier in the request phase. It should be relatively
straightforward to implement.