Dies Koper wrote on 03/10/2010 04:21 PM:
>> What happens to log files? Do all instances log to the same log
>> file? I assume not. I assume they each have their own log file
>> in the <instance>/logs directory and the agent/logs directory is
>> used for logs for the node agent itself. Correct?
>
> Yes.
> Also, a directory with the name of the instance will be created in the
> log directory to store the transaction logs. (Which are not
> human-readable, they contain data on on-going transactions, to be able
> to clean them up if the instance suddenly dies in the middle of a global
> transaction). These shouldn't be sync'ed either.
Yes, the log directory shouldn't be sync'ed.
Which log directory are the transaction logs stored in?
Hopefully the <instance>/logs directory. Why would it
need to create an <instance>/logs/<instance> directory
for the transaction logs?
>> If the instances do their own synchronization, it seems that we can't
>> support having a single config directory per node, shared by all the
>> instances. The config data will have to be moved under the
>> <instance>/config directory. Anyone see any problems with that?
>
> As it is in V2, right?
No, as I understand it (and I'm not an expert in this, I'm going on
what others have told me), the config data is in the node agent's
config directory, not the instance's config directory. I'm proposing
a change to that.
>> We need to allow both an instance and the DAS to run on the same machine
>> using the same software installation image. That means the instance
>> can't keep its data under the domains directory. With no node agent,
>> does it make sense to keep the data under the nodeagents directory?
>> We could use a fake node agent name such as "default" or "none" for
>> this case.
>>
>> With a single v2 software installation, multiple node agents supporting
>> instances in multiple domains can be created. With no actual node
>> agents, how can we support the same? Seems like we need multiple
>> "fake" node agent directories. Should the names of these directories
>> be derived from the domain name or the DAS address, at least by default,
>> so that users don't need to choose the name of a directory that doesn't
>> matter to them?
>
> As we discussed yesterday, there is only one node agent per domain per
> machine. So I thought, why not make the name exactly the same as the
> domain name?
Because the domain name is not unique. Multiple domains can have a
domain named "domain1".
>> Alternatively, we could enhance the create-node-agent command to have
>> a "--fake" option, or something like that, so that we could use it to
>> create the directory structure above. Then a "create-local-instance"
>> command (or "create-instance --local") could be used to configure the
>> instance-specific part.
>>
>> I'm leaning towards this latter approach.
>
> If the former approach is to rename the directories, why not both?
> I never really understood the reason to let users name the node agent.
You can name the node agent, but you're not required to. It just defaults
to the host name.
> The create-node-agent command (with the fake option) could be run as
> part of 'ant -f setup-cluster.xml' or whatever mechanism we'll have in
> GFv3.1 to enable cluster mode.
We shouldn't need any such command. It should always be prepared for
you to create a cluster.