[infinispan-dev] Logger lookup performance

[David M. Lloyd]

On 08/08/2011 05:06 AM, Sanne Grinovero wrote:
> Hi all,
> raising some attention here on Pete's suggestions;
> over the weekend we fixed this bootstrap performance issue ISPN-1315,
> which I had given the name "Reduce number of Logger instances being
> created",
> after looking at this profiler screenshot:
> http://community.jboss.org/servlet/JiveServlet/showImage/2-619820-16841/hotspot.jpg
>
> Now Pete suggested that the JBoss Logging should return the same
> instance; TBH from the screenshot all we know is that
> LogFactory.getLog(Class) invocations are taking 280 seconds, so I
> guess they are actually returning always the same instance, but and
> likely the time is spent in finding it. I'll rename the issue.

JBoss Logging definitely *tries* to return the same logger but there's 
no guarantee that it can (depending on the log backend).  In any case 
though, it's definitely better to use static final loggers as you say; 
this advice applies to every log framework I know of, not just JBoss 
Logging.

If you _really_ want covariant loggers (which generally I disapprove of, 
btw) then use a protected method which returns a static final logger. 
Note that most of the time the reason people use covariant loggers is so 
that the category name can reflect the instance class name.  However, I 
recommend against this for various reasons.  If the user wants the 
logging class name they can add it to the format.  If the user wants the 
instance class name, it should be part of the log message.  The logger 
category should be named to reflect the task being performed, not a 
class name.

> IMO we should stick with "static final" loggers, unless there are
> special reasons (and then please add a comment to clarify it was an
> intentional choice).

Right on.

-- 
- DML