Add log4j2 annotation processor so packages is not required

Currently it’s necessary to pass the packages attributes via <Configuration packages="io.prometheus.client.log4j2"> in log4j configuration to discover the appender. That means relying on step 5 of https://logging.apache.org/log4j/2.x/manual/plugins.html for appender discovery (analyzing packages and annotations at startup).

It would seem less error prone and more efficient to add the log4j annotation processor to the Maven compiler plugin, so it builds the metadata META-INF/org/apache/logging/log4j/core/config/plugins/Log4j2Plugins.dat at build time (step1) ?

[fstab]

Thanks for looking into this. So it looks like (2) is not a problem.

However, (1) sounds like this would break existing builds if users create uber-JARs but didn’t configure the log4j-transformer plugin. I imagine fixing the build will be non-trivial in some cases. Many frameworks like Spring inherit the Maven shade plugin from a parent POM, and it is not straightforward where to put the log4j-transformer plugin configuration.

[rquinio1A]

Good points, here’s what I’ve found:

1. For uber-jar we’re using this transformer to merge the Log4j2Plugins.dat fields: https://github.com/edwgiz/maven-shaded-log4j-transformer. It’s calling log4j-core APIs to do the merge and dump the resulting file.
2. log4j2 itself is iterating on all the classloader resources and merging plugins in memory: https://github.com/apache/logging-log4j2/blob/f72100df0decc9bda96b4d769822c4e48b2848fc/log4j-core/src/main/java/org/apache/logging/log4j/core/config/plugins/util/PluginRegistry.java#L162