We discovered a huge performance difference when initializing Log4J with deployed jar packages compared to using it directly with normal class files.
As it seems,
is way slower than its counterpart
and since this method gets called a few hundred times at initialization in this project, initialization of Log4J took more than 2 minutes with jar packages, compared to a few seconds with class files.
I just can tell that
org.apache.logging.log4j.core.config.plugins.PluginManager#collectPlugins() gets called VERY often via
org.apache.logging.log4j.core.config.BaseConfiguration#getPluginManager, which gets called by
And collectPlugins() is remapped to collectPlugins(boolean preLoad, final String pkgs) with collectPlugins(true, null);. I'm unsure what this preLoad is for, since it gets overridden with false and apparently no kind of caching is used. In this method, resolver.findInPackage(test, pkg); takes a whole lot of time searching within the jars, as described in the beginning.
The only way to work around this was for us to extract the custom appender in a sperate ant build target, and with this in a seperate jar, to minimize the size of the jar, reducing the search time.
Can you cache things and maybe reduce the calls to collectPlugins()? I don't see the probability of changing plugins during initialization as described in
LOG4J2-175 or the comment in org.apache.logging.log4j.core.config.BaseConfiguration#getPluginManager.
Introducing a variable for the plugin manager within the for loop in org.apache.logging.log4j.core.config.XMLConfiguration#constructHierarchy might be a good idea!