Re:Killing a process and all of its descendants
by pas on Sunday August 04, 2019 @08:03AM
(#20605975)
I mean MySQL did that with libaio (which needs O_DIRECT anyway), Oracle on Linux did that and some with ASM (their own filesystem basically), also Ceph with BlueStore (LevelDB as a FS).
Because general purpose filesystems and I/O susbsystems are just that, general purpose.
PostgreSQL regularly woes about the problems of the Linux I/O and fs APIs, but ... to my knowledge they did not do much else. (Other than working around the problems in userspace, and taking the performance hit.)
If you want I/O that doesn't throw you data away, use the perfectly fine buffered I/O on let's say XFS. Fast, safe, sane. If you need more juice, add NVMe, and/or MD-RAID / LVM. And eventually you need to scale out to multiple nodes anyway, and then the performance of the cluster consistency subsystem will be the weak point.
Because general purpose filesystems and I/O susbsystems are just that, general purpose.
PostgreSQL regularly woes about the problems of the Linux I/O and fs APIs, but ... to my knowledge they did not do much else. (Other than working around the problems in userspace, and taking the performance hit.)
If you want I/O that doesn't throw you data away, use the perfectly fine buffered I/O on let's say XFS. Fast, safe, sane. If you need more juice, add NVMe, and/or MD-RAID / LVM. And eventually you need to scale out to multiple nodes anyway, and then the performance of the cluster consistency subsystem will be the weak point.