On Nov 5, 2008, at 5:12 PM, Annette Jäkel wrote:
Am 30.10.2008 15:34 Uhr schrieb "Annette Jäkel" unter
<[EMAIL PROTECTED]>:
Am 30.10.2008 14:27 Uhr schrieb "Lars Marowsky-Bree" unter <[EMAIL PROTECTED]
>:
On 2008-10-30T12:40:43, Annette Jäkel <[EMAIL PROTECTED]>
wrote:
But my problem is that specific order of the nfsserver resource
after all
filesystem resources. This works fine until one of the file
system crashes.
Then nfsserver stops and the whole system doesnt work. Thats not
what I
want. If a file system crashes, for example because physical
device was
broken, the remaining devices and file systems should work and
exported to
clients. So maybe the nfsserver should start without any order?
But I'm not
shure that this is a good idea. I also dont want to run NFS, if
there is no
device it can manage.
Ah, OK, so the interleaving is not your highest priority.
I think rsc_order/_colocation constraints with score=0 achieve
exactly
what you need.
I think, I have to learn more about individual scores. Until now I
use
heartbeats default scores, never set a score manually. Meanwhile I
read the
linux-HA score basics docs. I understand your proposal this way:
score=0
mean, that nothing occurs because no score is changed. So if a
filesystem
resource fails, nfsserver resource should still run. Thats fine.
But whats
with the starting order of the resources? Does this occur in the
right order
if score of order=0? And what with the condition, that mounted
filesystems
and nfsserver MUST run on the same node? Does'nt this mean that
colocation
score should be INFINITY?
BTW: All order constraints in my current cib.xml are written
without a
score. Regarding to score documentation score then defaults to 0. All
colocation constraints defaults to INFINITY.
I will test the case on a testcluster with explicit setting of
score=0.
Meanwhile i tested my resource definitions for NFS Service and
corresponding
orders/colocations between resources for mounting filesystems and
starting
nfsserver. I tested:
1) order nfsserver after filesystems without explicit score setting
and
without any colocation definition
2) order nfsserver after filesystems with score=0 and without any
colocation
definition
3) stopping a filesystem resource and run a rcheartbeat stop/start
results:
- nfsserver started after filesystems
- regardless of order scores: if a filesystem resource was stopped,
nfsserver was still running resp. started at rcheartbeat start
4) order nfsserver after filesystems without explicit score setting
and with
a colocation from nfsserver to filesystems, score=INFINITY
result: stopping any of the filesystem resources also stops nfsserver
resource
4a) same constraints, stop a filesystem resource, rcheartbeat stop/
start
result: nfsserver stopped before heartbeat shutted down and dont
restart
after rcheartbeat start
5) order nfsserver after filesystems without explicit score setting
and with
a colocation from nfsserver to filesystems, score=0
results:
- if a filesystem resource was stopped, nfsserver was still running
- after stopping a filesystem resource and rcheartbeat stop/start,
nfsserver
resource started
Until now I thought I have a problem with the orders, but now I
think its
the problem of colocations. I have new questions:
1) is the behaviour of resources with regard to orders and
colocations the
same in case of stopping a resource as of failing of a resource
(resp.: was
my test set complete?)
Normally yes, but its possible to construct situations for which it
isn't (by using non-symmetrical ordering constraints).
2) Is there any sense for a colocation with score=0?
not really, but its not an either/or proposition.
you can also have 0 < score < INFINITY, which is "advisory" but not
mandatory.
Isn't it a hint that I also can remove this colocation?
right
discussion:
My intention for my current productive colocations between filesystem
resources and nfsserver resource is: nfsserver should always run on
the node
with the devices. But maybe I should remove colocations from
nfsserver to
all single filesystem resources and set a colocation only from
nfsserver to
the resource setting the MD Devices (but keep the orders of starting
nfsserver after filesystem resources!!)? My idea is: then I have the
right
order of resources and if md-devices fails, nfsserver stops - and
this is
ok.
But if a single filesystem resource fails, nfsserver is still running
like in my tests.
This is possible even with mandatory colocation... you just have to
get the constraint right.
<rsc_colocation from="X" to="Y" score="INFINITY"/>
is not the same as
<rsc_colocation from="Y" to="X" score="INFINITY"/>
The first version will allow Y to remain active when X stops, the
second will not.
Does that help?
Any opinion?
Best regards,
Annette
BTW: Now I will play with groups for all the filesystem resources.
Meanwhile I read within Michael Schwartzkopff's Linux-HA book. He
presented
an example NFS service with DRBD devices and if I understand all
setup steps
right, there is really no order from any resource to nfsserver. I
dont
understand this.
That would be wrong.
Within this context I think about setting up a group of all
filesystem
resources, setting attribute "ordered" to "false" and define a
single
resource of nfsserver ordered after the group - but dont wait for
the
complete start of the group, but starts if any resource of the
group is
running.
Well, that is the interleave part, which is not necessarily
required for
correctness, but for speeding up the start. That is not currently
possible.
I'm also not sure if it is indeed what you want; you want to have
tried
starting everything before bringing online "nfsserver", I think,
so that
they are all present to be exported - but not not start nfsserver if
just one of them has failed to start.
Think you're right. But does'nt this sounds like a scenario for a
on_fail=ignore|block for the start and monitor operation of every
filesystem
resource?
Regards,
Annette
In another thread (HA-NFS strategic question) Xinwei Hu suggested
not to
manage nfs but the step of exporting the devices, but seems
theres no RA for
this specific task. Until now I write /etc/exports by hand and call
"exportfs -a" from outside heartbeat, if I add or delete a
filesystem.
Yes, this is currently missing I think. Even Xinwei's nfsserver RA
is
"global" and doesn't support managing individual exports. It would
be
great if this was added.
Regards,
Lars
_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems
_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems
_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems
