Hi Mathias,
a looked at your patch and I think there are two bugs:
1) You set in_recovering of a woker which is in error_state to JK_TRUE even if
you don't select him for handling the request. This should be moved into the
next if clause when the worker is choosen.
2) When a local/main worker goes down and was set to error_state it would never
be asked again. Because if the local/main worker was in error and the time for a
new test has come you set rc only to this worker if it is no local worker and
this will never be true. And because of this and 1) in_recovering == JK_TRUE, so
it is dead for ever.
Thats why I have a more complicated if-statement in my patch. When I check if
the worker is in error state and this is true, I have to check if I can take it
anyway. Because I can't reach the else part of the if anymore.
My test environment is like in the explanation of my patch:
I got two nodes with one apache and tomcat on each node. So I have only one
local_worker per node and the mode is reject. When I shutdown tomcat I got an
error. But after starting up tomcat again it is still unreachable.
Bernd
Mathias Herberts wrote:
> Here is my patch to the load balancing code. I have tested the jk1
> part but not the jk2. The behavior should be identical though.
>
> The concept of local worker is introduced. Local workers form a subset
> of the balanced workers, those workers handle the requests with no
> session information or requests with session info but whose worker has
> failed.
>
> The list of local workers is specified in the workers.properties files
> as follows:
>
> worker.lb.balanced_workers=w1,w2,w3,w4,w5,w6
> worker.lb.local_workers=w4,w6
>
> Internally the module will rearrange the workers' list so that the
> local workers appear first in the list. For the example above the list
> will be rearranged as:
>
> w4,w6,w3,w1,w5,w2
>
> When a request comes in it is either part of a session or not. If the
> request is part of a session then the worker having created the
> session is selected (according to the session route appearing in the
> session id). If this worker has failed, the request is considered
> sessionless. For requests without a session, the workers' list is
> scanned from the beginning, selecting the worker not in error with the
> lowest lb_value. When the last local worker is reached, two
> alternatives exist depending on the value of the fault_action
> property. If fault_action is set to reject then if a worker was
> selected it is returned as the most suitable worker for the request,
> if no worker was selected so far, meaning all local workers are in
> error, a null value is returned as the most suitable worker, meaning
> the request is in error. If fault_action is set to balance, the
> selection processs continues.
>
> For jk1, if fault_action is set to balance, non local workers in error
> can be selected to handle the request. In jk2 in error workers are
> never selected.
>
> The fault_action is specified as follows:
>
> worker.lb.fault_action={reject|balance}
>
> With my patch, the lb_value or lb_factor needs not to have a special
> value to handle local workers. Thus lb_value is modified at each
> request handled by the worker.
>
>
> Feedback is welcome, testing of jk2 needs to be done.
>
> Mathias.
>
[...]
--
Dipl.-Inform. Bernd Koecke
UNIX-Entwicklung
Schlund+Partner AG
Fon: +49-721-91374-0
E-Mail: [EMAIL PROTECTED]
--
To unsubscribe, e-mail: <mailto:[EMAIL PROTECTED]>
For additional commands, e-mail: <mailto:[EMAIL PROTECTED]>
