[rabbitmq-discuss] Delay between minority detected and stopped server

Thu Aug 15 08:24:32 BST 2013

Malte,
On 15/08/13 08:00, Malte Schirmacher wrote:
> On 14/08/13 22:50, Matthias Radestock wrote:
>> There is *always* scope for that, since publishing is inherently
>> asynchronous, and so is detecting of cluster partitioning.
>
> Yet there *must not* be scope for this.

It's unavoidable. See below.

> There is more then a minute delay between the node noticing it's in
> minority and actually shutting down the node. Furthermore there is
> nothing here to be explained simply by saying message delivery is
> asynchonous as delivering messages was startet AFTER the node logged
> that its in minority.

The node logs that it is in a minority. Then it shuts down rabbit. That
takes time. Now, we can quibble about *how much* time, but it will
always be a non-zero amount. And during that time the node will still 
happily accept connections and route messages. Plus of course minority 
detection isn't instantaneous either, i.e. there is a time window during 
which the node technically in a minority but hasn't detected that yet.

Matthias.