On 12/5/12 12:06 PM, Galder Zamarreño wrote:
> On Dec 4, 2012, at 11:52 AM, Bela Ban <[email protected]> wrote:
>
>> If a node A connects to B and B connects to A at the exact same time
>> (and there wasn't any existing connection between the 2 nodes, then one
>> of the 2 will 'win' and the other one will close its connection. The
>> message to be sent is then lost.
>>
>> This is corrected by one of the upper layers, e.g. UNICAST retransmits
>> the message until it gets an ack. Re-sending a message will then create
>> a new connection, if the existing one was closed / removed.
>>
>> However, with UNICAST2, if a given message was the last message and no
>> further messages are sent, then only UNICAST2's stability messages will
>> detect that the other node is missing the last message sent. Stability
>> is triggered every 60 seconds by default, so unless that property was
>> changed, or stability was triggered programmatically, that last (lost)
>> message won't get retransmitted for 60 seconds.
> ^ Isn't that a default too high?

Assuming that the size-based stable messages are the norm, then 
time-based only kicks in as a second line of defense. With 
https://issues.jboss.org/browse/JGRP-1548 in place, this becomes even 
less important, so I want to leave it high as it does generate some 
traffic when set too small.


> Seems to me the scenario explained could happen relatively easily if two 
> nodes are started simultaneously.

No, the startup won't trigger concurrent connections, as only the joiner 
connects to the coordinator and the coordinator reuses the same 
connection to send the JOIN-RSP back.

It is the rebalancing process that triggers this in certain cases; the 
forwarding of state transfer requests to different owners.

> We no longer ask users to stagger their startups?

Because concurrent startup works, and not having to stagger startup is 
simpler.

-- 
Bela Ban, JGroups lead (http://www.jgroups.org)

_______________________________________________
infinispan-dev mailing list
[email protected]
https://lists.jboss.org/mailman/listinfo/infinispan-dev

Reply via email to