Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Monday 03 December 2007 10:59:34 Dominik Klein wrote: > Hi > > I read the recent thread "Suggestion to prevent split brain situation" > and this link about drbd resource fencing. > http://lists.linbit.com/pipermail/drbd-dev/2006-April/000341.html > > I ran a little test. I unplugged the DRBD link and the secondary was > still upToDate. I tried "drbdadm primary" on the secondary and it was > refused. After that, the resource was marked outdated. I believe you are mistaking parallelism for causality. :-) Mind you, I am totally shooting from the hip here as you didn't give us any logs or config files to look at. Specifically, I don't know what fencing policy you are using, whether you have redundant Heartbeat links, and whether you have set up dopd correctly. But I am led to believe that what really happened was this: The resource was marked Outdated *before* you attempted to make it Secondary. It just so happened by chance that it took the Primary a few seconds to outdate the resource on the Secondary, and you only noticed the completion of that process when you attempted "drbdadm primary" on the outdated node. > But now say the primary node fails, not just the DRBD link. How can the > secondary node become primary when it is marked outdated after having > lost communication to the primary node? Counter-question: How would a failed node, while no longer there, be able to outdate the surviving node? > Doesn't this prevent failover to work? No. :-) > What did I miss? I cannot imagine this is it. Could it be that you simply missed the fact that resource fencing is an active operation, where the Primary actively outdates its Secondary via a communication path outside of DRBD, and that such an operation doesn't complete instantaneously? Cheers, Florian -- : Florian G. Haas : LINBIT Information Technologies GmbH : Vivenotgasse 48, A-1120 Vienna, Austria