You need to make a custom build of Flume which uses log4j2.
You configure log4j2 to use a rolling file appender that rolls every minute (or whatever the latency is that you desire) to a spooling directory.
You configure Flume to use a SpoolingDirectorySource against that spooling directory.
You can't use a direct Flume appender (such as what's in log4j2) to log Flume because you will get into deadlock.
You can't use log4j1 with a rolling file appender because it has a concurrency defect which means it may write new messages to an old file and the SpoolingDirectorySource then fails.
I can't remember if I tried the Log4j appender from Flume with this setup. That appender does not have many ways to configure it and I think it will cause you problems if the subsequent agent you're trying to talk to is down.
Another approach might be to patch log4j1 and fix that concurrency defect (there's a variable that needs to be made volatile)
(Yes, setting this up is a little frustrating!)
