If I was implementing this I would take a custom JUL level and map it to the appropriate predefined JUL level. That would then map to a Log4j level.
There's actually a bit of an interesting challenge in converting from a custom level in JUL to Log4j. JUL allows you to use any integer value possible (not just non-negative ones). Also, their progression of level values goes in reverse of ours. Thus, any level above 1000 (Level.SEVERE in JUL) would need to be squeezed into the range of 1 to 99! Plus, Integer.MAX_VALUE indicates StandardLevel.ALL, but Level.OFF in JUL. Then there'd be the other way around, too.
Darn! That makes things tricky indeed...
1. Full auto: We could have some mapping logic that converts the custom JUL int level to a log4j int that is between the mapped built-in levels. (TBD: how to avoid collisions if multiple custom levels are defined between built-in levels?)
2. Semi-auto: we define an interface that converts JUL levels to Log4j levels. We provide a default impl for the built-in levels. Users need to provide their own impl (or extend ours?) if they have custom JUL levels. (TBD: how does our default impl handle undefined custom JUL levels?)
3. Config only: this depends on https://issues.apache.org/jira/browse/LOG4J2-589
Custom log4j levels are defined in configuration. The log4j config file is loaded first, so the JUL bridge can convert custom levels using the name only. It can completely ignore the JUL int level.
4. Easiest: we (initially) don't support custom JUL levels. Unknown levels are converted to some ad hoc log4j level. Let's say, INFO, but we can decide to use any level.
As to those fields, I think we can probably drop them. LogRecord dynamically calculates them from the Throwable stacktrace if necessary. We do it faster.
Phew!
What about the logp, entering, exiting, and throwing methods which all take a source class name and a source method name? Just ignore them?
My take would be to drop the seqNo and threadID integer, and for level, check if its a built-in JUL level which can be translated to a built-in log4j level. If it's not a built-in JUL level we can do a log4j Level.forName() call to create that custom level in log4j as well.
Thoughts?
Sent from my iPhone
I'm actually thinking of some sort of LogRecordMessage or similar which takes a useful subset of LogRecord.
I've got ranges in place to map to standard levels, but custom level support is currently done through the MDC. Should I use a MapMessage instead? Make a new Message type just for log4j-jul? There's metadata in some of these Logger methods that I'd like to include, but if the MDC isn't the best way to do that, then I'd prefer another way. I noticed that pax-logging does this for every log event to include some metadata about the OSGi bundle that made the log call, so I kept up the style.
As to the static field, yes, I noticed that, too. It's only for a sequence number, and we have our own (better) way of doing that with on-demand sequencing (and using the AtomicXxx classes indeed) anyways.
Fro a performance point of view, it would be great if we could avoid creating LogRecord instances. Not just from a GC perspective, but in java6 the LogRecord constructor synchronizes on a static variable(!): big bottleneck. This is improved (using AtomicXxx) in java7.
Also would great if we can avoid using the ThreadContext MDC for every log event. (Its copy-on-write design is not a good match for this usage...)
Would there be a way to map custom JUL log levels to custom Log4j levels?
Sent from my iPhone
Actually, now that I look at it, I can just use an inner class with ExtendedLoggerWrapper to get at those protected methods I mentioned. I mean, that appears to be the point of it! Let me see if it does everything I needed.
Now that I'm looking at this, what's the point of all the methods that take a FQCN instead of having just the ones in ExtendedLogger? I'm not sure why we didn't just use a field in AbstractLogger in the first place.
I'm making some changes to log4j-jul to reduce redundant time spent constructing a LogRecord that I don't even want to use most of the time. However, the ExtendedLogger interface (which I need to use at the very least so that I can set the fqcn to java.util.logging.Logger) only provides a single version of logMessage (unlike AbstractLogger which has a bunch), and several methods like catching(), throwing(), etc., all depend on protected methods in AbstractLogger that I'd rather not re-implement. It would be nice if I could just call the Logger methods I need, but they all get called with the wrong fqcn.
Can we use a non-static final field that contains the fqcn? If I could, I'd extend AbstractLogger myself, but I already have to extend the JUL Logger class (should have been an interface, grrr). Thus, I can't rely on AbstractLogger being the source of all these method calls. Unlike the other adapters, JUL provides more various logger calls than we even have, and I don't think ExtendedLogger was written with this scenario in mind.
I don't think this should be too large an impact of a change. I'm going to push up a proposal, but feel free to veto it or offer some suggestions!
--
--
--
--
--
--
--