[ClusterLabs] Antw: Re: Issue with DB2 HADR cluster

[Andrei Borzenkov]

On Wed, Apr 3, 2019 at 10:14 AM Ulrich Windl
<Ulrich.Windl at rz.uni-regensburg.de> wrote:
>
> >>> Digimer <lists at alteeve.ca> schrieb am 02.04.2019 um 19:49 in Nachricht
> <6c6302f4-844b-240d-8d0e-727dddf36950 at alteeve.ca>:
>
> [...]
> > It's worth noting that SBD fencing is "better than nothing", but slow.
> > IPMI and/or PDU fencing completes a lot faster.
>
> I'm surprised: Once sbd writes the fence command, it usually takes less than 3 seconds until the victim is dead. If you power off a server, the PDU still may have one or two seconds "power reserve", so the host may not be down immediately. Besides of that power-cycles are additional stress for the hardware...
>
> So maybe you want to explain why and how much faster IPMI and PDU fencing are.
>

We are not talking about death pill when both servers are in normal
state. We are talking about node failure, in which case second node
must wait for sufficient time to assume other node is really dead or
at least committed suicide. This time must also account for possible
storage timeouts including path failover so it cannot be too low,
otherwise sbd starts to commit suicide on every storage hiccup.