ext3 theoretically supports barriers when used with simple volumes. In practice, and for database purposes, they are not functional enough to help. The problem is that fsync calls are not correctly translated into write barrier in a form what will always flush the drive's cache. It's just not something built into ext3 in the right form, and if you are using Linux software RAID or the Logical Volume Manager (LVM) with ext3, barriers will not be available at all anyway.
What actually happens on ext3 when you execute fsync is that all buffered data on that filesystem gets flushed out in order to satisfy that call. You read that right cached data goes out every time fsync occurs. This is another reason why putting ...
