activateOptions pattern not desirable

[Curt Arnold <ca...@apache.org>]

I obviously haven't reviewed your last commit, but you did appear to 
tweak with activateOptions in a few classes and mention appender 
lifecycle which brought the topic to mind.

A very common pattern in log4j is to have a class that operates in a 
configuration mode early in its lifetime and then after a call to 
activate or activateOptions is ready for use.  Unfortunately this 
pattern implies objects that are not inherently immutable.  See 
"Effective Java", item 13 for the benefits of immutable classes.  Since 
the objects using the activateOptions pattern are subject to change 
early in the object's lifetime, it is only developer etiquette and 
self-preservation that enforces the object does not change during use.  
It would be better (in a perfect world) if distinct objects were used 
for the mutable configuration and immutable run-time behavior.

I don't have a path forward to move to greater use of immutable objects 
in log4j.  However, I think it should be seriously considered on any 
substantial rework.  When I finally get to my long standing date with 
LoggingEvent, it is likely to come out immutable.

When looking at appenders that change state over time, such as a 
database appender that can go into a failsafe mode if it can't connect, 
it would be preferable that these modes be represented by distinct 
immutable objects.  So you would have an object that forever 
represented a working database and an object that forever represented 
the failover mode and the failover transition was accomplished by 
switching these objects out instead of changing an object that has 
potentially been shared.

One pattern that might allow a transition is to have 
activateOptions-like methods return an object pointer which is used 
from then on.  Mutable objects could just return "this" since they 
don't change object identity when activating.  However, if one wanted 
to have an immutable implementation, the initial instance could be of a 
"Configurator" or "Builder" variant and the value returned from 
activateOptions would be immutable.  For example, an immutable 
FileAppender could be implemented with two classes: an immutable 
FileAppender and a FileAppenderBuilder, which allows setFile and the 
like, and would return a FileAppender on the call to activateOptions.


---------------------------------------------------------------------
To unsubscribe, e-mail: log4j-dev-unsubscribe@logging.apache.org
For additional commands, e-mail: log4j-dev-help@logging.apache.org

Re: activateOptions pattern not desirable

[Curt Arnold <ca...@apache.org>]

On Feb 18, 2005, at 1:30 PM, Ceki Gülcü wrote:

>
> I don't think Appenders can be completely immutable. For one, you can 
> invoke the close() method on an appender. You can also add remove 
> filters or change the threshold of an active appender. The only 
> appender which should be immutable is a closed appender.
>
> It does not make much sense to me to harden appenders when one can 
> easily yank out any appender from a logger and replace it with a 
> different one...
>
>

I was suggesting that if a substantial part of the framework is being 
refactored, trying to make the objects immutable should be considered 
as part of the refactoring.  Maybe Appenders aren't a good fit with the 
pattern, maybe they are but it just isn't obvious.  I wasn't suggesting 
that we try to refactor them now.

However, the state changes that you mentioned could be represented by a 
series of immutable objects.  The advantage is that you would not need 
to synchronize the state changes only the swap of the new appender with 
the old appender.  For example, if I'm removing one filter and adding 
another filter and I'm changing the object in place, the appender with 
just the first filter removed could be visible to other threads for a 
period of time.  With immutable objects, I could do the state changes 
in isolation and then swap out the new appender for the old one as an 
atomic operation.  It also can reduce the number of code points that 
you have to worry about synchronization.  With immutable objects, you 
only have to synchronize the replacement of one object with another.  
With mutable objects, you have to synchronize every method that might 
be affected by the contents of a mutable field.


---------------------------------------------------------------------
To unsubscribe, e-mail: log4j-dev-unsubscribe@logging.apache.org
For additional commands, e-mail: log4j-dev-help@logging.apache.org

Re: activateOptions pattern not desirable

[Ceki Gülcü <ce...@qos.ch>]

I don't think Appenders can be completely immutable. For one, you can 
invoke the close() method on an appender. You can also add remove filters 
or change the threshold of an active appender. The only appender which 
should be immutable is a closed appender.

It does not make much sense to me to harden appenders when one can easily 
yank out any appender from a logger and replace it with a different one...

At 07:03 PM 2/18/2005, Curt Arnold wrote:
>I obviously haven't reviewed your last commit, but you did appear to tweak 
>with activateOptions in a few classes and mention appender lifecycle which 
>brought the topic to mind.
>
>A very common pattern in log4j is to have a class that operates in a 
>configuration mode early in its lifetime and then after a call to activate 
>or activateOptions is ready for use.  Unfortunately this pattern implies 
>objects that are not inherently immutable.  See "Effective Java", item 13 
>for the benefits of immutable classes.  Since the objects using the 
>activateOptions pattern are subject to change early in the object's 
>lifetime, it is only developer etiquette and self-preservation that 
>enforces the object does not change during use.
>It would be better (in a perfect world) if distinct objects were used for 
>the mutable configuration and immutable run-time behavior.
>
>I don't have a path forward to move to greater use of immutable objects in 
>log4j.  However, I think it should be seriously considered on any 
>substantial rework.  When I finally get to my long standing date with 
>LoggingEvent, it is likely to come out immutable.
>
>When looking at appenders that change state over time, such as a database 
>appender that can go into a failsafe mode if it can't connect, it would be 
>preferable that these modes be represented by distinct immutable 
>objects.  So you would have an object that forever represented a working 
>database and an object that forever represented the failover mode and the 
>failover transition was accomplished by switching these objects out 
>instead of changing an object that has potentially been shared.
>
>One pattern that might allow a transition is to have activateOptions-like 
>methods return an object pointer which is used from then on.  Mutable 
>objects could just return "this" since they don't change object identity 
>when activating.  However, if one wanted to have an immutable 
>implementation, the initial instance could be of a "Configurator" or 
>"Builder" variant and the value returned from activateOptions would be 
>immutable.  For example, an immutable FileAppender could be implemented 
>with two classes: an immutable FileAppender and a FileAppenderBuilder, 
>which allows setFile and the like, and would return a FileAppender on the 
>call to activateOptions.
>
>
>---------------------------------------------------------------------
>To unsubscribe, e-mail: log4j-dev-unsubscribe@logging.apache.org
>For additional commands, e-mail: log4j-dev-help@logging.apache.org

-- 
Ceki Gülcü

   The complete log4j manual: http://www.qos.ch/log4j/



---------------------------------------------------------------------
To unsubscribe, e-mail: log4j-dev-unsubscribe@logging.apache.org
For additional commands, e-mail: log4j-dev-help@logging.apache.org