On Mon, May 2, 2011 at 3:51 PM, Lars Ellenberg
<[email protected]> wrote:
> On Mon, May 02, 2011 at 01:20:16PM +0200, Andrew Beekhof wrote:
>> On Mon, May 2, 2011 at 8:27 AM, Ulrich Windl
>> <[email protected]> wrote:
>> >>>> Andrew Beekhof <[email protected]> schrieb am 29.04.2011 um 09:31 in 
>> >>>> Nachricht
>> > <[email protected]>:
>> >> On Fri, Apr 29, 2011 at 9:27 AM, Dominik Klein <[email protected]> 
>> >> wrote:
>> >> > It waits $dampen before changes are pushed to the cib. So that
>> >> > eventually occuring icmp hickups do not produce an unintended failover.
>> >> >
>> >> > At least that's my understanding.
>> >>
>> >> correcto
>> >
>> > Hi!
>> >
>> > Strange: So the update is basically just delayed by that amount of
>> > time? I see no advantage: If you put a bad value to the CIB
>> > immediately or after some delay, the value won't get better by that.
>> > "Damping" siggests some filtering to me, but you are saying your are
>> > not filtering the values, but just delaying them. Right?
>>
>> Only the "current" value is written.
>> So the cluster will tolerate "minor" outages provided they last for
>> less than the dampen interval and the monitor frequency is high
>> enough.
>
> Still, we may get a spurious failover in this case:
>
> reachability:
>   +++++++++_____++++++++++++_____________________________
> Node A monitoring intervals:
>        +    -    +    +    +    -    -    -    -    -
> Node B monitoring intervals:
>     +    +    -    +    +    -    -    -    -    -
> "dampening" interval:         |---------|
>
> Note how the "dampening" helps to ignore the first network "glitch".
>
> But for the "permanent" network problem, we may get spurious failover:

Then your dampen setting is too short or interval too long :-)

>
> One dampening interval after node B notices loss of reachability,
> it will trigger a PE run, potentially moving things from B to
> A, because on A, the reachability (in the CIB) is still ok.
>
> Shortly thereafter, the dampening interval on A also expires, and the
> CIB will be updated with "A cannot reach out there either".
>
> Any resource migrations triggered by "B cannot reach out there"
> are now recognized as spurious.
>
> Question is, how could we avoid them?
>
> "ipfail" used to "ask the peer", wait for the peer to notice the new
> situation as well, and only then trigger actions.
>
> We could possibly store a short history of values, and actually do some
> "filtering" arithmetic with them.  Not sure if this should be done
> inside or outside of the CIB. Probably outside.

Yes, outside :-)

One of these attrd needs a rewrite :-(
_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems

Reply via email to