For the last week or so I've been testing and observing when hung jobs occur. And the answer is pretty clear.
With everything running 9.6.3, NO hung jobs, whether connecting to the DB cluster via HAproxy or direct to the local node. I can upgrade EVERYTHING BUT the Director to 9.6.5, including the one Fedora client that was still running 9.4.4, and still no hung jobs, with or without HAproxy. Or I can upgrade ONLY the Director to 9.6.5, with *everything* else running 9.6.3, and immediately get hung jobs, whether connecting directly or via HAproxy, but more common if using HAproxy. This problem is very clearly tied to the 9.6.5 Director as a regression from 9.6.3, and there are strong hints that it has to do with how the Director is talking to the database. I don't see anything in the 9.6.5 release notes that looks relevant, but my impression is that little Director code changed from 9.6.3 to 9.6.5. I do not know C++ and my C chops are pretty rusty at this point (and I was never a professional C programmer in the first place), but I've diffed 9.6.3/9.6.5 source and gone through the diffs line by line, and not spotted any change that looks obviously relevant to me. I don't know what else is needed to reproduce it. It might occur only with MariaDB 10.4, or only with attribute spooling disabled (necessary because the way attribute spooling works is not compatible with Galera clusters, because when the attributes are unspooled any batch size set appears to be ignored). At this point having confirmed that it occurs only with Director 9.6.5 — i.e, Director 9.6.5 alone is both necessary *and* sufficient — I'm now going to re-upgrade the Director to 9.6.5 and try again to identify exactly what is going wrong, but so far I really haven't identified anything conclusive. -- Phil Stracchino Babylon Communications ph...@caerllewys.net p...@co.ordinate.org Landline: +1.603.293.8485 Mobile: +1.603.998.6958 _______________________________________________ Bacula-devel mailing list Bacula-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/bacula-devel
