Question about replication, split-brains and unclean failover

[egoitz at sarenet.es]

Hi!,


I was writing some code for automating the server fail-over and was trying to see, how should or could I handle not run log files from sync_client. Well when in a clean shutdown, it’s pretty easy to know how to manage because the replication is up-to-date… so almost no problem there, it’s pretty fast... The problem comes in an unclean shutdown where some delay exists.

I sometimes suffer about unclean shutdowns, the way are described here https://www.fastmail.com/help/technical/architecture.html "Unclean failover”. There sais too, some improvements where going to be done (or that were committed perhaps) to the replication in order to avoid them. But, by the way replication is handled (as I have seen in the source code about locks, modseq checks and so) and the way Cyrus writes replication logs for rolling replication later, the possibility of replying the logs from an actual slave (just failed over), for covering a split-brain I assume is not carried out at least nowadays?. Perhaps am I wrong?. I ask this just, for confirming and avoid having wrong ideas… 

If it’s undone or at least partially undone, I would love really doing something… although unfortunately due to our high work load... I can’t say when I could have some time for it… but I’ll try my bests...



Cheers!