[ovirt-users] HA - Fencing not working when host with engine gets shutdown

Martin Perina mperina at redhat.com
Thu Sep 24 11:24:23 UTC 2015



----- Original Message -----
> From: "Michael Hölzl" <mh at ins.jku.at>
> To: "Martin Perina" <mperina at redhat.com>, "Eli Mesika" <emesika at redhat.com>
> Cc: "Doron Fediuck" <dfediuck at redhat.com>, users at ovirt.org
> Sent: Thursday, September 24, 2015 12:35:13 PM
> Subject: Re: [ovirt-users] HA - Fencing not working when host with engine gets shutdown
> 
> Hi,
> 
> thanks for the detailed answer! In principle, I understand the issue
> now. However, I can not fully follow the argument that this is a corner
> case. In a smaller or medium sized company, I would assume that such a
> setup, consisting of two machine with a hosted engine, is not uncommon.
> Especially as there is documentation online which describes how to
> deploy this setup. Does that mean that hosted engines are in general not
> recommended?
> 
> I am also wondering why the fencing could not be triggered by the hosted
> engine after the DisableFenceAtStartupInSec timeout? In the events log
> of the engine I keep on getting the message "Host hosted_engine_2 is not
> responding. It will stay in Connecting state for a grace period of 120
> seconds and after that an attempt to fence the host will be issued.",
> which would indicate that the engine is actually trying to fence the non
> responsive host.

Unfortunately this is a bit misleading message, it's shown every time that
we start handling network exception for the host and it's fired before
the logic which manages to start/skip fencing process (this misleading
message is fixed in 3.6). But in current logic we really execute fencing
only when host status is about to change from Connecting to NonResponsive
and this happens only for the 1st time when we are still in
DisableFenceAtStartupInSec interval. During all other attempts the host is
already in status Non Responsive, so fencing is skipped.

> 
> On 09/24/2015 11:50 AM, Martin Perina wrote:
> >
> > ----- Original Message -----
> >> From: "Eli Mesika" <emesika at redhat.com>
> >> To: "Martin Perina" <mperina at redhat.com>, "Doron Fediuck"
> >> <dfediuck at redhat.com>
> >> Cc: "Michael Hölzl" <mh at ins.jku.at>, users at ovirt.org
> >> Sent: Thursday, September 24, 2015 11:38:39 AM
> >> Subject: Re: [ovirt-users] HA - Fencing not working when host with engine
> >> gets shutdown
> >>
> >>
> >>
> >> ----- Original Message -----
> >>> From: "Martin Perina" <mperina at redhat.com>
> >>> To: "Michael Hölzl" <mh at ins.jku.at>
> >>> Cc: users at ovirt.org
> >>> Sent: Thursday, September 24, 2015 11:02:21 AM
> >>> Subject: Re: [ovirt-users] HA - Fencing not working when host with engine
> >>> gets shutdown
> >>>
> >>> Hi,
> >>>
> >>> sorry for the late response, but you hit a "corner case" :-(
> >>>
> >>> Let me start explain you a few things first:
> >>>
> >>> After startup of engine there's an internval during which fencing is
> >>> disabled. It's called DisableFenceAtStartupInSec and by default it's
> >>> set to 5 minutes. It can be changed using
> >>>
> >>>    engine-config -s DisableFenceAtStartupInSec
> >>>
> >>> but please do that with caution.
> >>>
> >>> Why do we have such timeout? It's a prevention of fencing storm, which
> >>> could happen in during power issues in whole DC: when both engine and
> >>> hosts are started, for huge hosts it may take a lot of time until become
> >>> up and VDSM start to communicate with engine. So usually engine is
> >>> started
> >>> first and without this interval engine will start fencing for hosts which
> >>> are just starting ...
> >>>
> >>> Another thing: if we cannot properly fence the host, we cannot determine
> >>> if there's not just communication issue between engine and host, so we
> >>> cannot restart HA VMs on another host. The only thing we can do is to
> >>> offer "Mark host as rebooted" manual option to administrator. If
> >>> administrator execution this option, we try to restart HA VMs on
> >>> different
> >>> host ASAP, because admin took the responsibility of validation that VMs
> >>> are really not running.
> >>>
> >>>
> >>> When engine is started, following actions related to fencing are taken:
> >>>
> >>> 1. Get status of all hosts from DB and schedule Non Responding Treatment
> >>>    after DisableFenceAtStartupInSec timeout is passed
> >>>
> >>> 2. Try to communicate with all host and refresh their status
> >>>
> >>>
> >>> If some host become Non Resposive during DisableFenceAtStartupInSec
> >>> interval
> >>> we skip fencing and administator will see message in Events tab that host
> >>> is Non Responsive, but fencing is disabled due to startup interval. So
> >>> administrator have to take care of such host manually.
> >>>
> >>>
> >>> Now what happened in your case:
> >>>
> >>>  1. Hosted engine VM is running on host1 with other VMs
> >>>  2. Status of host1 and host2 is Up
> >>>  3. You kill/shutdown host1 -> hosted engine VM is also shut down -> no
> >>>  engine
> >>>     is running to detect issue with host1 and change its status to Non
> >>>     Responsive
> >>>  4. In the meantime hosted engine VM is started on host2 -> it will read
> >>>  host
> >>>     status from DB, but all hosts are up -> it will try to communicate
> >>>     with
> >>>     host1,
> >>>     but it's unreachable -> so it changes host1 status Non Responsive and
> >>>     starts
> >>>     Non Responsive Treatment for host1 -> Non Responsive Treatment is
> >>>     aborted
> >>>     because engine is still in DisableFenceAtStartupInSec
> >>>
> >>>
> >>> So in normal deployment (without hosted engine) admin is notified that
> >>> host,
> >>> where engine is running, crashed and was rebooted, so he has to take a
> >>> look
> >>> and do manual steps if needed.
> >>>
> >>> In hosted engine deployment it's an issue because hosted engine VM can be
> >>> restart
> >>> on different host also in other cases then crashes (for example if host
> >>> is
> >>> overloaded hosted engine can stop hosted engine VM and restart it on
> >>> different
> >>> host, but this shouldn't happen too often).
> >>>
> >>> At the moment the only solution for this is manual: let administrator to
> >>> be
> >>> notified that host engine VM is restarted on different host, so
> >>> administrator
> >>> can check manually what was the cause for this restart and execute manual
> >>> steps
> >>> if needed.
> >>>
> >>> So to summarize: at the moment I don't see any reliable automatic
> >>> solution
> >>> for this :-( and fencing storm prevention is more important. But feel
> >>> free
> >>> to
> >>> create
> >>> a bug for this issue, maybe we can think of at least some improvement for
> >>> this use
> >>> case.
> >> Thanks for the detailed explanation Martin
> >> Really a corner case, lets see if we got more inputs on that from other
> >> users
> >> Maybe when hosted engine VM is restarted on another node we can ask for
> >> the
> >> reason and act accordingly
> >> Doron, with current implementation, is the reason for hosted engine VM
> >> restart stored anywhere ?
> > I have already discussed this with Martin Sivak and hosted engine doesn't
> > touch engine db at all. We discussed this possible solution with Martin,
> > which we could do in master and maybe in 3.6 if agreed:
> >
> >  1. Just after start of engine we can read from the db name of the host
> >     which hosted engine VM is running on and store it somewhere in memory
> >     for Non Responding Treatment
> >
> >  2. As a part of Non Responding Treatment we can some hosted engine
> >     specific logic:
> >       IF we are running as hosted engine AND
> >          we are inside DisableFenceAtStartupInSec internal AND
> >          non responsive host is the host stored above in step 1. AND
> >          hosted engine VM is running on different host
> >       THEN
> >          execute fencing for non responsive host even when we are
> >          inside DisableFenceAtStartupInSec internal
> >
> > But it can cause unnecessary fence for the case that whole datacenter
> > recovers from power failure.
> >
> >>>
> >>> Thanks
> >>>
> >>> Martin Perina
> >>>
> >>> ----- Original Message -----
> >>>> From: "Michael Hölzl" <mh at ins.jku.at>
> >>>> To: "Martin Perina" <mperina at redhat.com>
> >>>> Cc: users at ovirt.org
> >>>> Sent: Monday, September 21, 2015 4:47:06 PM
> >>>> Subject: Re: [ovirt-users] HA - Fencing not working when host with
> >>>> engine
> >>>> gets shutdown
> >>>>
> >>>> Hi,
> >>>>
> >>>> The whole engine.log including the shutdown time (was performed around
> >>>> 9:19)
> >>>> http://pastebin.com/cdY9uTkJ
> >>>>
> >>>> vdsm.log of host01 (the host which kept on running and took over the
> >>>> engine) split into 3 uploads (limit of 512 kB of pastebin):
> >>>> 1 : http://pastebin.com/dr9jNTek
> >>>> 2 : http://pastebin.com/cuyHL6ne
> >>>> 3 : http://pastebin.com/7x2ZQy1y
> >>>>
> >>>> Michael
> >>>>
> >>>> On 09/21/2015 03:00 PM, Martin Perina wrote:
> >>>>> Hi,
> >>>>>
> >>>>> could you please post whole engine.log (from the time which you turned
> >>>>> off
> >>>>> the host with engine VM) and also vdsm.log from both hosts?
> >>>>>
> >>>>> Thanks
> >>>>>
> >>>>> Martin Perina
> >>>>>
> >>>>> ----- Original Message -----
> >>>>>> From: "Michael Hölzl" <mh at ins.jku.at>
> >>>>>> To: users at ovirt.org
> >>>>>> Sent: Monday, September 21, 2015 10:27:08 AM
> >>>>>> Subject: [ovirt-users] HA - Fencing not working when host with engine
> >>>>>> gets
> >>>>>> 	shutdown
> >>>>>>
> >>>>>> Hi all,
> >>>>>>
> >>>>>> we are trying to setup an ovirt environment with two hosts, both
> >>>>>> connected to a ISCSI storage device, a hosted engine and power
> >>>>>> management configured over ILO. So far it seems to work fine in our
> >>>>>> testing setup and starting/stopping VMs works smoothly with proper
> >>>>>> scheduling between those hosts. So we wanted to test HA for the VMs
> >>>>>> now
> >>>>>> and started to manually shutdown a host while there are still VMs
> >>>>>> running on that machine (to simulate power failure or a kernel panic).
> >>>>>> The expected outcome was that all machines were HA is enabled, are
> >>>>>> booted again. This works if the machine with the failure does not have
> >>>>>> the engine running. If the machine with the hosted engine VM gets
> >>>>>> shutdown, the host gets in the "Not Responsive state" and all VMs end
> >>>>>> up
> >>>>>> in an unkown state. However, the engine itself starts correctly on the
> >>>>>> second host and it seems like it tries to fence the other host (as
> >>>>>> expected) - Events which we get in the open virtualization manager:
> >>>>>> 1. Host hosted_engine_2 is non responsive
> >>>>>> 2. Host hosted_engine_1 from cluster Default was chosen as a proxy to
> >>>>>> execute Status command on Host hosted_engine_2.
> >>>>>> 3. Host hosted_engine_2 became non responsive. It has no power
> >>>>>> management configured. Please check the host status, manually reboot
> >>>>>> it,
> >>>>>> and click "Confirm Host Has Been Rebooted"
> >>>>>> 4. Host hosted_engine_2 is not responding. It will stay in Connecting
> >>>>>> state for a grace period of 124 seconds and after that an attempt to
> >>>>>> fence the host will be issued.
> >>>>>>
> >>>>>> Event 4 is continuously coming every 3 minutes. Complete engine.log
> >>>>>> file
> >>>>>> during engine boot up: http://pastebin.com/D6xS3Wfy
> >>>>>> So the host detects the machine is not responding and wants to fence
> >>>>>> it.
> >>>>>> But although the host has power management configured over ILO, the
> >>>>>> engine thinks that it is not. As a result the second host does not get
> >>>>>> fenced and VMs are not migrated to the running machine.
> >>>>>> In the log files there are also a lot of time out exception. But I
> >>>>>> guess
> >>>>>> that this is because the host cannot connect to the other machine.
> >>>>>>
> >>>>>> Did anybody face similar problems with HA? Or any clue what the
> >>>>>> problem
> >>>>>> might be?
> >>>>>>
> >>>>>> Thanks,
> >>>>>> Michael
> >>>>>>
> >>>>>>
> >>>>>> ----
> >>>>>> ovirt version: 3.5.4
> >>>>>> Hosted engine VM OS: Cent OS 6.5
> >>>>>> Host Machines OS: Cent OS 7
> >>>>>>
> >>>>>> P.S. We also have to note that we had problems with the command
> >>>>>> fence_ipmilan at the beginning. We were receiving the message "Unable
> >>>>>> to
> >>>>>> obtain correct plug status or plug is not available," whenever the
> >>>>>> command fence_ipmilan was called. However, the command fence_ilo4
> >>>>>> worked. So we use a simple script for fence_ipmilan now that calls
> >>>>>> fence_ilo4 and passes the arguments.
> >>>>>> _______________________________________________
> >>>>>> Users mailing list
> >>>>>> Users at ovirt.org
> >>>>>> http://lists.ovirt.org/mailman/listinfo/users
> >>>>>>
> >>> _______________________________________________
> >>> Users mailing list
> >>> Users at ovirt.org
> >>> http://lists.ovirt.org/mailman/listinfo/users
> >>>
> 



More information about the Users mailing list