[Pgcluster-general] Failover weakness

[a.mitani at sra-europe.com]

Hi Tom,

It is depend on the OS whether it detect unpluged cable as network error.
The application on linux can not detect it from socket error.

My multi-node test is runing on VMware.
Therefore, I can not unpluged from it.

Regards,
-----------------
At.Mitani

> Did you try to unplug ethernet cable of node-A instead of bringing down
> node-A services?
> Your test (with the exception of recovery start) works for me, except
> that if I unplug node-A from the network, things doesn't work.
> Also, I guess that the results will be pretty the same if you cut the
> power of node-A instead of unplugging it from the network.
>
> Regards,
> --
> Tom;
>
> a.mitani at sra-europe.com
>> Hi Tom,
>>
>>
>>> Again, I have nodes A and B both with cluster and replication server
>>> (synchronized). When node A or B gets disconnected from the network
>>> (whether the active replicator is in A or B), attempting to execute a
>>> write query on the connected DB hangs the query for, say, infinite
>>> time.
>>> The conclusion, failover doesn't work as one might expect because I
>>> have
>>> N fail-points for N nodes :( .
>>>
>>> Is this normal, or just happened to me?
>>>
>>
>> In my test environment, the Cluster A is recovered as normally.
>> test was done as follow.
>>
>> (1) start cluster DBs
>> cluster db A started in node A.
>> cluster db B started in node B.
>>
>> (2) start replication servers
>> replicator A started in node A.
>> replicator B started in node B.
>>
>> (3)create DB
>> send "createdb" command to cluster A.
>>
>> (4)stop node A
>> cluster db A and replicator A were downed.
>>
>> (5)run pgbench
>> run pgbench in cluster B.
>>
>> (6)re-start cluster db A
>> start cluster A with "-R" option in node A.
>> <<cluster db A was recovered as normally.>>
>>
>> (7)re-start replicator A
>> just start replicator A in node A
>>
>> (8)run pgbench again
>> run pgbench in cluster db B.
>> <<it was replicated in cluster db A as well.>>
>>
>> I tried same test in 3 nodes.
>> The result is same (stopped cluster db is recovered as normally).
>>
>> I believe that this recovery function is normal.
>> However, it seems not easy to find problem with present package.
>>
>> I'll send a patch for adding debug message to you.
>>
>>
>> At.Mitani
>>
>> _______________________________________________
>> Pgcluster-general mailing list
>> Pgcluster-general at pgfoundry.org
>> http://pgfoundry.org/mailman/listinfo/pgcluster-general
>>
>
> _______________________________________________
> Pgcluster-general mailing list
> Pgcluster-general at pgfoundry.org
> http://pgfoundry.org/mailman/listinfo/pgcluster-general
>