As is described in Flume-3342, renaming a file may cause data duplication. But moreover, if the original file was opened by flume before and not closed yet, flume reopens it while not freeing its original file handles. And then I delete the new file, but after the idle time, flume just closes the new file handle and forgets to close its original file handles, leading to a handle leak. It can be found by "lsof | grep "deleted".
This bug is triggered by using Log4j to roll log files. To avoid losing data when rolling, I use regex to include both current log file and old log files in the taildir path.