[Gluster-devel] Re: Unexpected behaviour when self-healing

Daniel Wirtz daniel at virtunity.com
Wed May 28 07:01:25 UTC 2008


Ok, I updated to 1.3.9 now but there is still some strange behaviour. I
tested the same case like mentioned before:

- I start up both servers with a fresh and clean datastorage and namespace.
- Then I create a file named "before" which is replicated correctly between
both servers.
- Now I killall -9 one of the two server nodes so simulate a crash.
- Afterwards I create a new file named "after" on the still working server
and remove the "before" file.
- When this took place I restart the crashed node (note: with 1.3.9 it does
not crash anymore the first time it starts).

Now, I see both files "before" and "after" on clients that are connected
either to the first or the second server.
However, I cannot open the "before" file on the not-crashed server node (I
deleted it from there before). This could be an expected behaviour, but both
files show up on both servers. However, the "after" file is located only in
the exported directory of the not-crashed server and the before file is
located only in the exported directory of the crashed server. There seems no
replication to be done to any of the two files. However, both servers
have namespace entries for both files (but the "before" file cannot be
opened on the not-crashed server as mentioned).

Some of the files located inside the underlying exported directories also
have a timestamp of 1970-01-01 01:00. This seems to be the case only on the
server, where the file was automatically replicated to but not on the
server where it originally was created (I am using ext3 as underlying
filesystem). Is this intended? However, the files that show up inside the
mounted directories got the correct timestamps.

-> As for now it seems that previously deleted files that exist on a
returning crashed node but were removed from the node that did not crash in
the meantime, show up on both nodes and are still accessible on the crashed
node but not on the not-crashed node. Also there seems no replication to
take place for both kinds of files - that are files created before the crash
but deleted afterwards on the remaining server and files that are created
between crash and restart on the not-crashed server.

Am I doing something wrong?

regards
Daniel



More information about the Gluster-devel mailing list