On 18/04/17 10:00 AM, Digimer wrote: > On 18/04/17 03:47 AM, Ulrich Windl wrote: >>>>> Digimer <li...@alteeve.ca> schrieb am 16.04.2017 um 20:17 in Nachricht >> <12cde13f-8bad-a2f1-6834-960ff3afc...@alteeve.ca>: >>> On 16/04/17 01:53 PM, Eric Robinson wrote: >>>> I was reading in "Clusters from Scratch" where Beekhof states, "Some would >> >>> argue that two-node clusters are always pointless, but that is an argument >>> for another time." Is there a page or thread where this argument has been >>> fleshed out? Most of my dozen clusters are 2 nodes. I hate to think they're >> >>> pointless. >>>> >>>> -- >>>> Eric Robinson >>> >>> There is a belief that you can't build a reliable cluster without >>> quorum. I am of the mind that you *can* build a very reliable 2-node >>> cluster. In fact, every cluster our company has deployed, going back >>> over five years, has been 2-node and have had exception uptimes. >>> >>> The confusion comes from the belief that quorum is required and stonith >>> is option. The reality is the opposite. I'll come back to this in a minute. >>> >>> In a two-node cluster, you have two concerns; >>> >>> 1. If communication between the nodes fail, but both nodes are alive, >>> how do you avoid a split brain? >> >> By killing one of the two parties. >> >>> >>> 2. If you have a two node cluster and enable cluster startup on boot, >>> how do you avoid a fence loop? >> >> I think the problem in the question is using "you" instead of "it" ;-) >> Pacemaker assumes all problems that cause STONITH will be solved by STONITH. >> That's not always true (e.g. configuration errors). Maybe a node's failcount >> should not be reset if the node was fenced. >> So you'll avoid a fencing loop, but might end in a state where no resources >> are running. IMHO I'd prefer that over a fencing loop. >> >>> >>> Many answer #1 by saying "you need a quorum node to break the tie". In >>> some cases, this works, but only when all nodes are behaving in a >>> predictable manner. >> >> All software relies on the fact that it behaves in a predictable manner, BTW. >> The problem is not "the predictable manner for all nodes", but the >> predictable >> manner for the cluster. >> >>> >>> Many answer #2 by saying "well, with three nodes, if a node boots and >>> can't talk to either other node, it is inquorate and won't do anything". >> >> "wan't do anything" is also wrong: I must go offline without killing others, >> preferrably. >> >>> This is a valid mechanism, but it is not the only one. >>> >>> So let me answer these from a 2-node perspective; >>> >>> 1. You use stonith and the faster node lives, the slower node dies. From >> >> Isn't there a possibility that both nodes shoot each other? Is there a >> guarantee that there will always be one faster node? >> >>> the moment of comms failure, the cluster blocks (needed with quorum, >>> too) and doesn't restore operation until the (slower) peer is in a known >>> state; Off. You can bias this by setting a fence delay against your >>> preferred node. So say node 1 is the node that normally hosts your >>> services, then you add 'delay="15"' to node 1's fence method. This tells >>> node 2 to wait 15 seconds before fencing node 1. If both nodes are >>> alive, node 2 will be fenced before the timer expires. >> >> Can only the DC issue fencing? >> >>> >>> 2. In Corosync v2+, there is a 'wait_for_all' option that tells a node >>> to not do anything until it is able to talk to the peer node. So in the >>> case of a fence after a comms break, the node that reboots will come up, >>> fail to reach the survivor node and do nothing more. Perfect. >> >> Does "do nothing more" mean continuously polling for other nodes? >> >>> >>> Now let me come back to quorum vs. stonith; >>> >>> Said simply; Quorum is a tool for when everything is working. Fencing is >>> a tool for when things go wrong. >> >> I'd say: Quorum is the tool to decide who'll be alive and who's going to die, >> and STONITH is the tool to make nodes die. If everything is working you need >> neither quorum nor STONITH. >> >>> >>> Lets assume that your cluster is working find, then for whatever reason, >>> node 1 hangs hard. At the time of the freeze, it was hosting a virtual >>> IP and an NFS service. Node 2 declares node 1 lost after a period of >>> time and decides it needs to take over; >> >> In case node 1 is DC, isn't a selection for a new DC coming first, and the >> new >> DC doing the STONITH? >> >> >>> >>> In the 3-node scenario, without stonith, node 2 reforms a cluster with >>> node 3 (quorum node), decides that it is quorate, starts its NFS server >>> and takes over the virtual IP. So far, so good... Until node 1 comes out >> >> Again if node 1 was DC, it's not that simple. >> >>> of its hang. At that moment, node 1 has no idea time has passed. It has >> >> You assume no fencing was done... >> >>> no reason to think "am I still quorate? Are my locks still valid?" It >>> just finishes whatever it was in the middle of doing and bam, >>> split-brain. At the least, you have two nodes claiming the same IP at >>> the same time. At worse, you had uncoordinated writes to shared storage >>> and you've corrupted your data. >> >> But that's no cluster; that's a mess ;-) >> >>> >>> In the 2-node scenario, with stonith, node 2 is always quorate, so after >>> declaring node 1 lost, it moves to fence node 1. Once node 1 is fenced, >>> *then* it starts NFS, takes over the virtual IP and restores services. >> >> So you compare "2 nodes + fencing" to "3 nodes without fencing"? >> >>> In this case, no split-brain is possible because node 1 has rebooted and >>> comes up with a fresh state (or it's on fire and never coming back anyway). >>> >>> This is why quorum is optional and stonith/fencing is not. >> >> You did not convince me how only one node has the ability to fence the other >> without a quorum: Wouldn't both nodes shoot at each other? (I quoted this so >> many times, but once again: In HP-UX Service Guard, a lock disk was used as a >> tie-breaker: Only one node suceeded to get the lock, and the other committed >> suicide (via kernel watchdog timeout)). >> >>> >>> Now, with this said, I won't say that 3+ node clusters are bad. They're >>> fine if they suit your use-case, but even with 3+ nodes you still must >>> use stonith. >>> >>> My *personal* arguments in favour of 2-node clusters over 3+ nodes is this; >> >> Again: You compare "2 nodes with fencing" to "3 nodes without fencing". My >> personal vote would be "3 nodes with fencing" if there is enough work for two >> nodes. >> >>> >>> A cluster is not beautiful when there is nothing left to add. It is >>> beautiful when there is nothing left to take away. >>> >>> In availability clustering, nothing should ever be more important than >>> availability, and availability is a product of simplicity. So in my >>> view, a 3-node cluster adds complexity that is avoidable, and so is >>> sub-optimal. >> >> IMHO: a valid cluster software works starting at 1 node, then per induction >> also for n+1 nodes. Complexity should grow only linear with the number of >> nodes. Of course you shouldn't add nodes just for the number of nodes, but >> for >> the actual need. >> >> Regards, >> Ulrich > > I was addressing the misconception that fencing was optional and quorum > was not. I wrote a longer reply as an article to follow up on this down > the thread.
As an addendum; I will say, with clarity, that *all* clusters need stonith, period. 3+ node without stonith is still a disaster waiting to happen. -- Digimer Papers and Projects: https://alteeve.com/w/ "I am, somehow, less interested in the weight and convolutions of Einstein’s brain than in the near certainty that people of equal talent have lived and died in cotton fields and sweatshops." - Stephen Jay Gould _______________________________________________ Users mailing list: Users@clusterlabs.org http://lists.clusterlabs.org/mailman/listinfo/users Project Home: http://www.clusterlabs.org Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf Bugs: http://bugs.clusterlabs.org
