[rabbitmq-discuss] Rabbitmq cluster design and HA

[Michael Klishin]

Unless you mostly do request-reply workloads, just use disk nodes.

MK

> On 01/05/2014, at 15:26, srikanth tns <srikanthtns at gmail.com> wrote:
> 
> Also , we are trying to use 2DISC 1RAM in the cluster . Is this good enough ? Also which node should be set as RAM , is there any preference to set only a particular node the cluster to be RAM ?
> 
> Thanks
> Srikanth
> 
> 
>> On Wed, Apr 30, 2014 at 1:48 PM, srikanth tns <srikanthtns at gmail.com> wrote:
>> Thanks Micheal , here is the settings we are doing on rabbitmq cluster
>> 
>> {rabbit, [{vm_memory_high_watermark_paging_ratio, 0.8},{vm_memory_high_watermark, 0.9},{disk_free_limit, 1000000000}]}
>> 
>> Apart from this ,is there any QoS value that we need to set ?
>> 
>> 
>>> On Tue, Apr 29, 2014 at 1:13 AM, Michael Klishin <mklishin at gopivotal.com> wrote:
>>> 
>>> 
>>> On 29 April 2014 at 07:22:03, srikanth tns (srikanthtns at gmail.com) wrote:
>>> > > Can you let us know what would be optimal settings on rabbitmq
>>> > cluster to prevent it from crashing and having reliable messaging?
>>> 
>>> It's really hard to make general recommendations without knowing what
>>> kind of workload you use.
>>> 
>>> > 1) sysctl configurations
>>> 
>>> nlimit > 20,000 or so (definitely can't be less than ~ 6K)
>>> 
>>> > 2) Disk limitations to setup on cluster . the default is 50MB
>>> 
>>> I'd use 500 MB or so if RabbitMQ is the primary thing that runs on the machine.
>>> 
>>> > 3) Memory to used , right now its default 40%
>>> 
>>> Can be up to 90%, again, depending on what else may run on the same machine.
>>> 
>>> 
>>> > 6) any disk space limitation on the hosts
>>> 
>>> Host is this different from 2) above?
>>> 
>>> > For HA , we are using ha-mode all to duplicate the queues across
>>> > all nodes in the cluster. Do you think its reliable during the
>>> > failover scenario ?During the failover it would like 100k queues
>>> > on nthe new master? Or should we use ha-mode = exactly/nodes ?
>>> 
>>> With 3 nodes ha-mode=all is a reasonable amount of duplication. If you are confident
>>> you will never lose more than 1 node at a time, use ha-mode=exactly with 2 nodes.
>>> Less duplication means better throughput.
>>> 
>>> As to how well new master election will work with 100K queues, try it. 
>>> --
>>> MK
>>> 
>>> Software Engineer, Pivotal/RabbitMQ
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.rabbitmq.com/pipermail/rabbitmq-discuss/attachments/20140501/891c8a7a/attachment.html>