[Linux-ha-dev] Re: confused about the bully election algorithm implementation in crmd

[Andrew Beekhof]

I've committed the following change which should prevent this from
happening again:
    http://hg.beekhof.net/lha/crm-stable/rev/e4a4f12e9c2d

On 12/27/06, home_king <home_king at 163.com> wrote:
> Hi, Andrew.
>
> It is hard to repeat the problem. Because the pitfall that I described
> only work under the edge conditions, such as DC has a high payload, or
> some crmd broadcasts NON-VOTE messages slowly.
>
> However, I indeed ever encountered that condtions.
>
> This pitfall is based on two facts:
> 1. VOTE may be launched in non-atomic manner
> 2. Once receiving a NON-VOTE from some node, no matter it's outdated
> or not, no matter the 'voted' hash is empty (because the lastest ELECTION
> is already over) that time or not, it will be recorded in the 'voted' hash.
>
> Of course, I also found there exists some casual remedy for the pitfall.
> 1. callback of CCM event will skip the obsolete items in 'voted'
> 2. g_hash_table_replace() will replace the old key with the new key
> (NO-VOTEDs
> which come from the same node), though their corresponding
> "current_election_id" are different.
>
> The exist documents is old, will you please provide (write) some
> fresh & detail documents (about the design & key implementation of crmd
> family)
> for the linux-ha-fans? After all, the code of crmd family (including cib,pe,
> te,ccm,lrm) take up the biggest part of the whole hb, I think.
>
> _______________________________________________________
> Linux-HA-Dev: Linux-HA-Dev at lists.linux-ha.org
> http://lists.linux-ha.org/mailman/listinfo/linux-ha-dev
> Home Page: http://linux-ha.org/
>