Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
/ 2007-02-17 00:58:37 +0000 \ Rick Rothstein: > Greg Freemyer <greg.freemyer at ...> writes: > > > > > On 2/15/07, Greg Freemyer <greg.freemyer at ...> wrote: > > <snip> > > > As to: > > > > > > "You can stack DRBD on top of md, md on top of DRBD is nonsense, however". as I'm the author of that remark... I think it is nonsense, because it adds too much complexity. say, you have a failure. network hickup comes first, then the currently active box goes down. when you have a network hickup, some of your component drbds have experienced connection loss before the others. now when the other node takes over, the raid 0 stripe set of drb devices has corrupted data, since some did still receive writes from the other node when some had already lost the connection. even more strange things will happen in an active active setup, when parts of the drbd's lose connection before the others: e.g. the drbd where the file system journal lies was still connected, but some other part in the data area lost connection... all you can expect from a recovery attempt is now: nonsense. my strong recommendation is still: don't do that. it will come after you and haunt you and get you sooner or later. > > > I still don't know why that is true in a active-passive setup. It > > > seems totally logical to me to stack md raid0 above drbd raid1 in a > > > active-passive setup, but that may be different thread. > > > > > I was thinking some more about this and I believe the issue is with > > ordered writes. > > > > IIRC, drbd provides 3 modes of operation related to ordered writes. > > By stacking a raid layer above drbd you break its ability to control > > ordered writes because no single instance of drbd sees the whole > > picture. I'm not a drbd expert, but if you look at the 3 modes you > > can probably see why it would be bad even in a active passive setting. this is true, too, but a non-issue as long as you use the strictly synchronous drbd protocol C. > > Greg > I'm thinking (hoping?), that because each separate DRBD will preserve the > correct write ordering of its chunk of the raid-0 stripe; the overall write > ordering of the entire stripe will also be preserved. > > I guess that I'm just going to have to dig into the combinitorial discussion > in the DRBD docs; and then, of course there's tons of testing to do. > > _Rick -- : Lars Ellenberg Tel +43-1-8178292-0 : : LINBIT Information Technologies GmbH Fax +43-1-8178292-82 : : Vivenotgasse 48, A-1120 Vienna/Europe http://www.linbit.com : __ please use the "List-Reply" function of your email client.