Stephen Willey
2006-Jun-18 11:18 UTC
[Lustre-discuss] How available does the MGS have to be?
Reading the mountconf pages, it states (7.4) that if the MGS server goes down, the only effect is that configuration changes are not propagated. It also says (7.3) that if servers are started before the MGS is up, they won''t receive configuration changes. What would be the proper procedure in the event that it went down, got brought back up, but you didn''t want to restart all your other nodes? Is there any method by which you can ''reconnect'' the servers(/clients?) to it? Also, 7.3 only mentions servers that are started before the MGS is brought up. Should we take this to mean that clients are not affected? Will they pick up changes if an MGS is later brought back online? Thanks, Stephen
Brent A Nelson
2006-Jun-18 11:18 UTC
[Lustre-discuss] How available does the MGS have to be?
If you can do a failover MGS, though, presumably this would not be a problem (unless you lost both MGS nodes or the failover didn''t start)? On Wed, 24 May 2006, Nathaniel Rutman wrote:> Stephen Willey wrote: > >> Reading the mountconf pages, it states (7.4) that if the MGS server goes >> down, the only effect is that configuration changes are not propagated. >> It also says (7.3) that if servers are started before the MGS is up, >> they won''t receive configuration changes. >> >> What would be the proper procedure in the event that it went down, got >> brought back up, but you didn''t want to restart all your other nodes? >> Is there any method by which you can ''reconnect'' the servers(/clients?) >> to it? >> > There is no such procedure in place yet. This needs to become an automated > function, but is quite > complicated and has not been completed. >> Also, 7.3 only mentions servers that are started before the MGS is >> brought up. Should we take this to mean that clients are not affected? >> Will they pick up changes if an MGS is later brought back online? >> > Same deal for clients - if they lose the connection with the MGS, they no > longer receive updates. > Clients can''t start if the MGS is down, so there is no "started before the > MGS" for them. > > > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss@clusterfs.com > https://mail.clusterfs.com/mailman/listinfo/lustre-discuss >
Nathaniel Rutman
2006-Jun-18 11:18 UTC
[Lustre-discuss] How available does the MGS have to be?
Stephen Willey wrote:>Reading the mountconf pages, it states (7.4) that if the MGS server goes >down, the only effect is that configuration changes are not propagated. > It also says (7.3) that if servers are started before the MGS is up, >they won''t receive configuration changes. > >What would be the proper procedure in the event that it went down, got >brought back up, but you didn''t want to restart all your other nodes? >Is there any method by which you can ''reconnect'' the servers(/clients?) >to it? > >There is no such procedure in place yet. This needs to become an automated function, but is quite complicated and has not been completed.>Also, 7.3 only mentions servers that are started before the MGS is >brought up. Should we take this to mean that clients are not affected? > Will they pick up changes if an MGS is later brought back online? > >Same deal for clients - if they lose the connection with the MGS, they no longer receive updates. Clients can''t start if the MGS is down, so there is no "started before the MGS" for them.
Nathaniel Rutman
2006-Jun-18 11:18 UTC
[Lustre-discuss] How available does the MGS have to be?
It is still a problem as of beta3 (basically, there is no mechanism to re-enqueue a "config" lock that was never taken). This is a planned feature addition. But as of now, for the cases listed below, the only way to get a configuration update on a server or client that has lost (or never took) a configuration lock is to restart that server or client. Brent A Nelson wrote:> If you can do a failover MGS, though, presumably this would not be a > problem (unless you lost both MGS nodes or the failover didn''t start)? > > On Wed, 24 May 2006, Nathaniel Rutman wrote: > >> Stephen Willey wrote: >> >>> Reading the mountconf pages, it states (7.4) that if the MGS server >>> goes >>> down, the only effect is that configuration changes are not propagated. >>> It also says (7.3) that if servers are started before the MGS is up, >>> they won''t receive configuration changes. >>> >>> What would be the proper procedure in the event that it went down, got >>> brought back up, but you didn''t want to restart all your other nodes? >>> Is there any method by which you can ''reconnect'' the servers(/clients?) >>> to it? >>> >> There is no such procedure in place yet. This needs to become an >> automated function, but is quite >> complicated and has not been completed. >> >>> Also, 7.3 only mentions servers that are started before the MGS is >>> brought up. Should we take this to mean that clients are not affected? >>> Will they pick up changes if an MGS is later brought back online? >>> >> Same deal for clients - if they lose the connection with the MGS, >> they no longer receive updates. >> Clients can''t start if the MGS is down, so there is no "started >> before the MGS" for them. >> >> >> _______________________________________________ >> Lustre-discuss mailing list >> Lustre-discuss@clusterfs.com >> https://mail.clusterfs.com/mailman/listinfo/lustre-discuss >> > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss@clusterfs.com > https://mail.clusterfs.com/mailman/listinfo/lustre-discuss