[mirror-admin] "file has vanished" again

[Carlos Carvalho]

Rob van Nieuwkerk (robn at berrymount.nl) wrote on 17 October 2009 00:13:
 >On Fri, 16 Oct 2009 14:05:34 -0500 (CDT)
 >Mike McGrath <mmcgrath at redhat.com> wrote:
 >
 >> On Fri, 16 Oct 2009, Chris Adams wrote:
 >> 
 >> > Once upon a time, Mike McGrath <mmcgrath at redhat.com> said:
 >> > > Do you see it with mirroring other distributions?
 >> >
 >> > The only other major thing I mirror is CPAN, and I don't see it there.
 >> >
 >> 
 >> It seems impossible to avoid this situation completely since you'll always
 >> get a file list at one point in time, and finish syncing at a later point
 >> in time.

As I said earlier today, this is not the case. The vanished messages I
posted were from a list-only run, where rsync just sends the receiver
the file list. This means that the files "vanished" between a
readdir and the stat.

 >Of course, that's a thing that one would expect to happen.
 >But the weird thing here is that this situation can STAY for many
 >hours (or even days !) on the same file(s) with different rsync runs.

Yes. And I'm syncing from sync.fedoraproject.org.

If it were the actimeo setting it'd have to be on the order of hours
or days. And then it'd affect many files, not just the few that
produce errors each time.

 >I've always suspected it to have something to do with (incorrectly)
 >caching things on load-balancing equipment or NFS-setup.  Or nodes
 >in a load-balancing cluster that are out of sync.  Or nodes that are
 >ping-ponging between old & new content because of incorrect syncing-loops.

The list-only single-connection failures exclude these.

 >Do you know if your rsync server is maybe adapted (for performance reasons)
 >to fetch it's file-list from a different location (so not every rsync run
 >forces a scan over the whole tree) ?

Then it'd get the names and also the attributes, and wouldn't fail on
list-only runs.

Do you have very different cache times for directories and files? If
cache times for directories were much bigger than for files it could
happen that the directory cache has become stale and when rsync stats
the files there's a mismatch. However I don't believe this, because
the diretory cache times would have to be on the order of hours or
larger.

I bet this wouldn't happen if the repository was not remotely
mounted...

Does the nfs server have special tricks for directories?

--