On 29/01/19 16:42, Michal Soltys wrote: > > With supposedly so many problematic disks - wouldn't filesystem > journaling completely fall apart if flushes were not working correctly > (regardless whether it's flush or fua) ? Or a flush sent from within > e.g. VM. Or anything relying on fsync(). Just watch out for anything relying on fsync. It will absolutely kill performance under many circumstances. I don't remember the gory details, but for some reason the ext3 defaults were "fsync-friendly". When ext4 came out, the defaults changed, and programs that relied on fsync brought systems to their knees. I don't know what if anything has changed since then. I'm interested in databases as well as raid, and the general consensus there seems to be "don't use fsync, the semantics and implementation vary so much that what works for one person will be a disaster for another. It's not worth the grief". (bit like the latest alloca() security hole - the *nix implementation is so bad that even allocating ONE byte could be enough to bring down the system or enable a security breach) Cheers, Wol
