On Tue, May 30, 2023 at 10:02 AM Yu Kuai <yukuai1@xxxxxxxxxxxxxxx> wrote: > > Hi, > > 在 2023/05/30 9:36, Guoqing Jiang 写道: > > > > > > On 5/29/23 16:40, Xiao Ni wrote: > >> It's the data which customers read that can be wrong. So it's a > >> dangerous thing. Because customers use the data directly. If it's true > >> it looks like there is a race between non aligned read and recovery > >> process. > > > > But did you get similar report from customer from the past years? I am not > > sure it is reasonable to expect the md5sum should be match under such > > conditions (multiple processes do write/read with recovery in progress). I didn't get a similar report these years besides this one. > > > > Maybe the test launched 32 processes in parallel is another factor, not > > sure > > it still happens with only 1 process which means only do read after write. > > May I ask if these processes write the same file with same offset? It's > insane if they do... If not, this cound be a problem. They write to different files. One process writes to its own file. > > > > > Anyway, I think io accounting is irrelevant. > > Yes, I can't figure out why io accounting can be involved. Thanks, I'll look more. Regards Xiao > > Thanks, > Kuai > > > > Thanks, > > Guoqing > > . > > >
