On Sep 22, 2011, at 10:16 PM, Curt Arnold wrote: > Been lurking and watching the thread, wanted to throw some comments out > without fully following the thread. Take everything with a grain of salt.
Why? Your input is just as valuable as everyone else's. > > log4j 1.2's AsyncAppender does a bit of this by calling all sorts of methods > to force the lazy initialization of the event prior to the event being queued > for processing by the worker thread. This can result in a lot of unnecessary > work on the calling thread copying things like the MDC that might but might > not be used in the eventual message formatting. For some reason I haven't gotten around to the AsyncAppender. I'm afraid some copying of data when generating the LogEvent is inevitable. > > Performing the rendering of the message on the calling thread is also not > desirable since things like formatting floating point numbers can be > surprisingly expensive. One advantage of the Message interface is that it can provide information that can help determine when it should be rendered. For example, in addition to the initialize method John proposed for the DeferredMessage it could also indicate when it should be deferred to. Of course, Messages never need to format themselves until getFormattedMessage is called. > > In some of my earlier design experiments, I was very fond of having the > formatter/layout object supporting a two phase approach where the layout had > an opportunity to "extract" from the logging event on the calling thread and > then process the extracted information later on the worker thread. The > extraction method would be responsible to assemble the information that it > needed into an arbitrary immutable object that would be passed blindly by the > framework to the formatting phase at the proper time. The implementation had > the choice of whether it wanted to render all the way to String on the > calling thread (say if it were cheap) on do the minimum to get immutable > objects in the calling thread and do all the heavy lifting on the worker > thread. I'm not sure I see the benefit of creating another object. You still need the LogEvent. Most of the work to insure the LogEvent can be passed to another thread really needs to be done anyway. Currently the constructor copies the MDC and NDC and a ThrowableProxy is constructed if a Throwable is present. About the only things that aren't being done then is the setting of the caller's StackTraceElement and setting the name of the current thread. It is a simple matter to have the AsyncAppender cause those to be set. FWIW, the LogEvent should be immutable from the perspective of anything using it. Since LogEvent is an interface and only exposes get methods this would be true, except that getContextMap and getContextStack don't say they return unmodifiable objects and they implementation isn't returning immutable objects. They should. > > Without getting the layout participating, you almost will be doing > unnecessary work on the calling thread. I don't necessarily disagree with this part. For example, the Converters that deal with the caller's location could call LogEvent.getSource during the first call to make sure the required information in the LogEvent is captured. Of course, they would call that same method during the rendering but that second call would be very inexpensive. > > I fleshed this approach out several years ago at > https://svn.apache.org/repos/asf/logging/sandbox/experimental/pattern-layout > if anyone wants to take a look for ideas. I'll take a look at this. Ralph --------------------------------------------------------------------- To unsubscribe, e-mail: log4j-dev-unsubscr...@logging.apache.org For additional commands, e-mail: log4j-dev-h...@logging.apache.org
