7:38 a.m.
On 12/16/2009 12:14 PM, Manik Surtani wrote:
On 16 Dec 2009, at 09:26, Galder Zamarreno wrote:
>
>
> On 12/15/2009 01:13 PM, Manik Surtani wrote:
>> A few comments:
>>
>> - Why do you have OpCode in your response header? Surely this is redundant? If
the client is sync, it knows what it sent. If it is async, it has a message ID.
>
> True, I fixed that.
>
>> - 'Not So Dumb' and 'Clever' response headers should be optional?
Surely this stuff is only sent when there is a topology change? We also may need some
extra info here - how does the back-end know to send this info? If a client hits
different back-end nodes, and there is a topology change, how does Node A decide that it
should not bother with topology info since the client already hit Node B after the topo
change and has the new topology map? Perhaps a TopologyVersion (== JGroups View ID)
should be sent back with any topo map, and the client would send it's current
TopologyVersion with every request (non-dumb clients only)? Could be a vlong...
>
> Yeah, as you and Mircea suggest, clients sending the view id could help
> better manage size of responses.
>
> Something to note here is that not all cluster view changes necessarily
> involve a client side topology view change, cos if you have N infinispan
> nodes in the cluster, you could have M running hot rod server where M<=
> N. So, Hot rod will need to figure out when there's been a view change
> in the cluster that involves the addition or removal of a hot rod
> running Infinispan instance (similar thing to what happens when a
> clustered EJB is deployed, the target list for that clustered EJB gets
> updated).
Yes.
> So, bearing in mind this, could we just use the JGroups view id for
> this? AFAIK, it's 0 based long but shouldn't cause problems with
> complete cluster restarts. If the whole cluster gets restarted, existing
> connections will be closed and at that point, clients could revert back
> to trying to connect one of their known hot rod servers and pass -1 as
> view id which means that I have no view, so the responding server would
> send back the hot rod cluster view.
Provided the client knows the servers have all been restarted.
I think a better approach is that the client sends its last known topology version
(vClient). The server node it connects to compares vClient with vServer (the version that
the servers are aware of). if vClient != vServer then resend topology map.
I added that already this morning to the request header :)
http://community.jboss.org/wiki/HotRodProtocol
Not-so-dumb and clever clients send their latest viewId to the server
and the server compares would compare it and if different, send the new
topology.
Note that in the response, I had added a "View change marker" field so
that the client knows whether a view change follows or not. The reason I
had done this is because I didn't think the client could simply compare
the vServer with the vClient and decide whether more needs to be read or
not.
You could have an scenario where a client starts up and sends 3 request
in pararell, all with vClient=-1 because no replies have been received
from hot rod server. When the server replied to those 3 req, it would
have sent for example, vServer=4. The first response to be processed
would have update the local vClient to 4, but the other two responses
would still contain the vServer=4. So, they would somehow need to read,
or discard the response.
And after writing this and thinking it again, the "View change marker"
is not be necessary. If vClient and vServer are equals, the client knows
the total length of the body so it can read it and discard it directly.
> I'll add this to the wiki.
>
>>
>> Cheers
>> Manik
>>
>>
>> On 14 Dec 2009, at 20:08, Galder Zamarreno wrote:
>>
>>> Hi all,
>>>
>>> Re: http://community.jboss.org/wiki/HotRodProtocol
>>>
>>> I've updated the wiki with the following stuff:
>>> - Renamed replaceIfEquals to replaceIfUnmodified
>>> - Added remove and removeIfUnmodified.
>>> - Added containsKey command.
>>> - Added getWithCas command so that cas value can be returned. I decided
>>> for a separate command rather than adding cas to get return because you
>>> don't always want cas to be returned. Having a separate command makes
>>> better use of network bandwith.
>>> - Added stats command. JMX attributes are basically accessible through
>>> this, including cache size.
>>> - Added error handling section and updated status codes.
>>>
>>> Note that Mircea added some interesting comments and I replied to them
>>> directly in the wiki.
>>>
>>> Still remaining to add:
>>> - Commands: putForExternalRead evict, clear, version, name and quite
>>> commands.
>>> - Passing flags.
>>>
>>> Regards,
>>>
>>> p.s. Updating this has been quite a struggle due to F12 + FF 3.5.5
>>> crashing at least 5 times, plus parts of the wiki dissapearing after
>>> publishing them!
>>> --
>>> Galder Zamarreño
>>> Sr. Software Engineer
>>> Infinispan, JBoss Cache
>>> _______________________________________________
>>> infinispan-dev mailing list
>>> infinispan-dev(a)lists.jboss.org
>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>
>> --
>> Manik Surtani
>> manik(a)jboss.org
>> Lead, Infinispan
>> Lead, JBoss Cache
>> http://www.infinispan.org
>> http://www.jbosscache.org
>>
>>
>>
>>
>>
>> _______________________________________________
>> infinispan-dev mailing list
>> infinispan-dev(a)lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>
> --
> Galder Zamarreño
> Sr. Software Engineer
> Infinispan, JBoss Cache
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev(a)lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev
--
Manik Surtani
manik(a)jboss.org
Lead, Infinispan
Lead, JBoss Cache
http://www.infinispan.org
http://www.jbosscache.org
_______________________________________________
infinispan-dev mailing list
infinispan-dev(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev
--
Galder Zamarreño
Sr. Software Engineer
Infinispan, JBoss Cache