Question about recovery in cluster envirnment
Hi , Suppose I have mulitple nfs-ganesha servers behind a load balancer. All nfs clients connect the load balancer, then balancer choose a nfs-ganesha server and redirect the request to it. When a nfs-ganesha (Ganesha-A)which has connection with a client restart in a time, the server enter grace period and read previous clients from backend. However, as client connection is broken and load balancer choose another nfs-ganesha sever(Ganesha-B). Even the client want to reclaim open and clock, Ganesha-B is not in grace period. In this scenario, how could client relaim the resource? Thanks, Marvin
Hi Marvin, On 4/2/19 3:08 PM, fanzi2009@hotmail.com wrote:
Hi , Suppose I have mulitple nfs-ganesha servers behind a load balancer. All nfs clients connect the load balancer, then balancer choose a nfs-ganesha server and redirect the request to it. When a nfs-ganesha (Ganesha-A)which has connection with a client restart in a time, the server enter grace period and read previous clients from backend. However, as client connection is broken and load balancer choose another nfs-ganesha sever(Ganesha-B). Even the client want to reclaim open and clock, Ganesha-B is not in grace period. In this scenario, how could client relaim the resource?
Thats right. A simple load-balancer will not do the trick. You shall need special handling to put entire cluster (all ganesha servers) in grace even if one server goes down or restarts. There is dbus command available to put a ganesha server in grace anytime while it is active. And you would need to configure your cluster manager to invoke that command for all the servers when needed (for eg., for pcs/pacemaker using pacemaker resource agents; for CTDB using event scripts) Thanks, Soumya
Thanks,
Marvin _______________________________________________ Devel mailing list -- devel@lists.nfs-ganesha.org To unsubscribe send an email to devel-leave@lists.nfs-ganesha.org
On 4/2/19 5:38 AM, fanzi2009@hotmail.com wrote:
Hi , Suppose I have mulitple nfs-ganesha servers behind a load balancer. All nfs clients connect the load balancer, then balancer choose a nfs-ganesha server and redirect the request to it. When a nfs-ganesha (Ganesha-A)which has connection with a client restart in a time, the server enter grace period and read previous clients from backend. However, as client connection is broken and load balancer choose another nfs-ganesha sever(Ganesha-B). Even the client want to reclaim open and clock, Ganesha-B is not in grace period. In this scenario, how could client relaim the resource? Thanks, Marvin
If you're going to have multiple Ganesha's serving the same clients, whether load balanced or with failover, you need to have a system to put the entire set of Ganeshas into Grace when any of them fails. In this case, I suspect the load balancer needs to know when a Ganesha fails, and block any new client connections and not move any clients between Ganeshas until all the Ganeshas have entered Grace. Basically, the load balancer moving a client from one Ganesha to another, for any reason, looks like a failover to NFS, and all the Ganeshas need to be in Grace before it can happen. NFS is not well designed for a clustered server setup, so it's a bit awkward to get it working well. Daniel
Thanks for Daniel and Soumya! Putting whole cluster into grace period, will it cause performance issue? Is it OK that we save all the session informations into backend. When ServerA crashed and client connect ServerB, ServerB can resume session from backend including open and locking resource.
On 4/2/19 9:07 AM, fanzi2009@hotmail.com wrote:
Thanks for Daniel and Soumya! Putting whole cluster into grace period, will it cause performance issue?
Is it OK that we save all the session informations into backend. When ServerA crashed and client connect ServerB, ServerB can resume session from backend including open and locking resource.
You have to do that, and you also have to put the cluster into Grace. When the client connects to the new server, it will re-acquire any state it had before, and so the new server needs to know what clients can do this. As I said, NFS is not a great clustered protocol. If you have multiple servers serving the same data, and any of them goes down, you need to put them all into Grace. This is to keep any *other* clients from getting locks that should belong to the clients that were connected to the down server. Yes, this will have performance issues, but it's the only way to avoid data corruption. Daniel
On Tue, Apr 2, 2019 at 9:21 AM Daniel Gryniewicz <dang@redhat.com> wrote:
On 4/2/19 9:07 AM, fanzi2009@hotmail.com wrote:
Thanks for Daniel and Soumya! Putting whole cluster into grace period, will it cause performance issue?
Is it OK that we save all the session informations into backend. When ServerA crashed and client connect ServerB, ServerB can resume session from backend including open and locking resource.
You have to do that, and you also have to put the cluster into Grace. When the client connects to the new server, it will re-acquire any state it had before, and so the new server needs to know what clients can do this.
As I said, NFS is not a great clustered protocol. If you have multiple servers serving the same data, and any of them goes down, you need to put them all into Grace. This is to keep any *other* clients from getting locks that should belong to the clients that were connected to the down server. Yes, this will have performance issues, but it's the only way to avoid data corruption.
I mostly agree with everything Dan said here. Clustered environments require extra care, and we don't currently have a way to migrate clients between different ganesha heads that are exporting the same clustered fs. I've done some experimentation with v4 migration here, but there's nothing in stock ganesha today for this. You can, however, avoid putting all of the other cluster nodes into the grace period if you know that none of the other cluster nodes can acquire state that will need to be reclaimed by a clients of the node that is restarting. We have some tentative plans to implement this for FSAL_CEPH + the RADOS recovery backends someday, but it requires support in the Ceph MDS and userland client libraries that does not yet exist. -- Jeff Layton <jlayton@poochiereds.net>
How about I persist all the session information into backend(rados for example)? If so, even though the client connect to another server, the server can reconstruct session from backend. Client don't need to re-create session and don't need to reclaim.
On Tue, Apr 2, 2019 at 9:21 AM Daniel Gryniewicz <dang(a)redhat.com> wrote:
I mostly agree with everything Dan said here. Clustered environments require extra care, and we don't currently have a way to migrate clients between different ganesha heads that are exporting the same clustered fs. I've done some experimentation with v4 migration here, but there's nothing in stock ganesha today for this.
You can, however, avoid putting all of the other cluster nodes into the grace period if you know that none of the other cluster nodes can acquire state that will need to be reclaimed by a clients of the node that is restarting. We have some tentative plans to implement this for FSAL_CEPH + the RADOS recovery backends someday, but it requires support in the Ceph MDS and userland client libraries that does not yet exist.
This might work, but keep in mind that *any* change to the client (or it's associated state) needs to be persisted to synchronously rados, which will introduce huge latency to most ops. This isn't even memory -> disk latency (100x - 1000x) but memory to network latency (10000x or higher). Daniel On 4/2/19 10:46 PM, fanzi2009@hotmail.com wrote:
How about I persist all the session information into backend(rados for example)? If so, even though the client connect to another server, the server can reconstruct session from backend. Client don't need to re-create session and don't need to reclaim.
On Tue, Apr 2, 2019 at 9:21 AM Daniel Gryniewicz <dang(a)redhat.com> wrote:
I mostly agree with everything Dan said here. Clustered environments require extra care, and we don't currently have a way to migrate clients between different ganesha heads that are exporting the same clustered fs. I've done some experimentation with v4 migration here, but there's nothing in stock ganesha today for this.
You can, however, avoid putting all of the other cluster nodes into the grace period if you know that none of the other cluster nodes can acquire state that will need to be reclaimed by a clients of the node that is restarting. We have some tentative plans to implement this for FSAL_CEPH + the RADOS recovery backends someday, but it requires support in the Ceph MDS and userland client libraries that does not yet exist.
Devel mailing list -- devel@lists.nfs-ganesha.org To unsubscribe send an email to devel-leave@lists.nfs-ganesha.org
Easier said than done. Bear in mind that all of the recovery backend stuff is entirely for the purpose of dealing with server restarts, so you really do have to be careful not to leave gaps in particular scenarios. Let's say you do decide to synchronously store open and lock records in a central RADOS-based database and both Server A and Server B are using it. Server A crashes, and the client decides to reconnect to Server B using the recorded clientid/session. Server B says "Oh, this session was previously held by Server A." Now what happens? You need a mechanism to transfer the CephFS state (opens, locks, caps, etc.) from Server A to Server B. Nothing like that exists today, but we do have some tentative plans to allow cephfs clients to reclaim state they previously held. In principle, that could be extended to allow "takeover" in some fashion. But wait...it gets worse! Suppose we have a 3 node ganesha cluster and some of Server A's clients decide to go to Server C instead? Now a simple takeover is not enough -- you need a way to split that state granularly. Couple all of this with the basic truism that failures in these sorts of architectures are often cascading. You need to deal with the possibility that any node could just die at any time, and decide how you're going to deal with that. A lot of the original ganesha recovery backend work had gaping holes in the "takeover" mechanisms where failures at an inopportune time could make it so no clients could recover anything. This is very much a non-trivial problem in my experience, but don't let me dissuade you if you have considered these scenarios and have thoughts on how to address it. -- Jeff Layton <jlayton@poochiereds.net> On Tue, Apr 2, 2019 at 10:46 PM <fanzi2009@hotmail.com> wrote:
How about I persist all the session information into backend(rados for example)? If so, even though the client connect to another server, the server can reconstruct session from backend. Client don't need to re-create session and don't need to reclaim.
On Tue, Apr 2, 2019 at 9:21 AM Daniel Gryniewicz <dang(a)redhat.com> wrote:
I mostly agree with everything Dan said here. Clustered environments require extra care, and we don't currently have a way to migrate clients between different ganesha heads that are exporting the same clustered fs. I've done some experimentation with v4 migration here, but there's nothing in stock ganesha today for this.
You can, however, avoid putting all of the other cluster nodes into the grace period if you know that none of the other cluster nodes can acquire state that will need to be reclaimed by a clients of the node that is restarting. We have some tentative plans to implement this for FSAL_CEPH + the RADOS recovery backends someday, but it requires support in the Ceph MDS and userland client libraries that does not yet exist.
Devel mailing list -- devel@lists.nfs-ganesha.org To unsubscribe send an email to devel-leave@lists.nfs-ganesha.org
One more thing to keep in mind: Getting this wrong is a data corruption bug, which is the worst possible bug you can have in a storage system, so you should be pretty sure you've gotten it right. Daniel On 4/5/19 9:20 AM, Jeff Layton wrote:
Easier said than done. Bear in mind that all of the recovery backend stuff is entirely for the purpose of dealing with server restarts, so you really do have to be careful not to leave gaps in particular scenarios.
Let's say you do decide to synchronously store open and lock records in a central RADOS-based database and both Server A and Server B are using it.
Server A crashes, and the client decides to reconnect to Server B using the recorded clientid/session. Server B says "Oh, this session was previously held by Server A." Now what happens?
You need a mechanism to transfer the CephFS state (opens, locks, caps, etc.) from Server A to Server B. Nothing like that exists today, but we do have some tentative plans to allow cephfs clients to reclaim state they previously held. In principle, that could be extended to allow "takeover" in some fashion.
But wait...it gets worse!
Suppose we have a 3 node ganesha cluster and some of Server A's clients decide to go to Server C instead? Now a simple takeover is not enough -- you need a way to split that state granularly.
Couple all of this with the basic truism that failures in these sorts of architectures are often cascading. You need to deal with the possibility that any node could just die at any time, and decide how you're going to deal with that. A lot of the original ganesha recovery backend work had gaping holes in the "takeover" mechanisms where failures at an inopportune time could make it so no clients could recover anything.
This is very much a non-trivial problem in my experience, but don't let me dissuade you if you have considered these scenarios and have thoughts on how to address it.
participants (4)
-
Daniel Gryniewicz -
fanzi2009＠hotmail.com -
Jeff Layton -
Soumya Koduri