tomcat-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From bugzi...@apache.org
Subject [Bug 54045] New: ReplicatedMap don't like TcpFailureDetector in static configuration
Date Tue, 23 Oct 2012 16:03:53 GMT
https://issues.apache.org/bugzilla/show_bug.cgi?id=54045

          Priority: P2
            Bug ID: 54045
          Assignee: dev@tomcat.apache.org
           Summary: ReplicatedMap don't like TcpFailureDetector in static
                    configuration
          Severity: major
    Classification: Unclassified
          Reporter: frederic.arnoud@gmail.com
          Hardware: PC
            Status: NEW
           Version: 7.0.32
         Component: Cluster
           Product: Tomcat 7

Tribes stack using:
* TcpPingInterceptor
* TcpFailureDetector
* MessageDispatchInterceptor
* StaticMembershipInterceptor
Do not work well in static cluster.

First side (ie one thread):
* call to TcpFailureDetector.heartbeat()
* call to checkMembers(false)
* call to performBasicCheck() in synchronized(membership)
* in performBasicCheck, for a missing static node:
*   add "missing" member to membership with membership.memberAlive(m)
*       HERE THE SECOND THREAD HAVE SOME TIME TO WORK
*   check it with memberAlive(m)
*   remove it since if it doesn't exist

Second side (ie another thread):
* some call to channel.getMembers() like what the done by AbstractReplicatedMap
* this call will call the TcpFailureDetector.getMembers()
* this one could return a wrong value since it can contains unavailable nodes

Note:
* synchronize on membership isn't use by TcpFailureDetector in getMember(),
getMembers(), hasMembers(), neither in Membership equivalent method (maybe
because it's too heavy to lock every thread while the TcpFailureDetector check
if node are alive).

It must not be an issue for AbstractReplicatedMap since with or without
TcpFailureDetector a node could disapear while replicated map try to use it.
But ReplicatedMap use always Channel.SEND_OPTIONS_DEFAULT where the value is
Channel.SEND_OPTIONS_USE_ACK. So a message sent to a missing node will fail
with an exception.

Personnaly I override TcpFailureDetector.heartbeat() to avoid
performBasicCheck() if I use a static configuration (TcpPingInterceptor call
performForcedCheck()).
But this doesn't fix ReplicatedMap issue.

Better fix could avoid adding missing member to membership list:
* Add a method like memberAlive(MemberImpl) to Membership without side effect
(add the member)
* in TcpFailureDetector.performBasicCheck(): check this new method before
adding the node


This doesn't fix the AbstractReplicatedMap issue which work always with
acknoledge from other nodes.


Same code for Tomcat 6.


best regards
F.Arnoud

-- 
You are receiving this mail because:
You are the assignee for the bug.

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@tomcat.apache.org
For additional commands, e-mail: dev-help@tomcat.apache.org


Mime
View raw message