何时在log4j配置中使用随机访问文件附加器与缓冲文件附加器

Question

I've been looking at optimizing my Log4J configuration to improve the throughput and reduce latency, and I'm trying to determine the best settings. I'm dealing with a system where the code must be bug-free, reliable, and system performance are the top priority in order from most important to least important.

It seems pretty clear to me that Asynchronous loggers is the clear winner in terms of best performance, and that makes complete sense. I don't understand the tradeoffs of Random Access File Appenders vs. Buffered File Appenders. I looked at the log4j website here but I didn't see any real downside to using Random Access File Appenders.

Could somebody please explain the differences, and explain when each should be used?

Answer 1

Points of interest:

• RandomAccessFileAppenders are always buffered. As of beta-9, the buffer size is 256 * 1024 bytes and not configurable. This will become configurable in the next release.

• FileAppenders can be configured with the bufferedIO attribute to be buffered or not. As of beta-9, the buffer size is 8 * 1024 bytes and not configurable. This will become configurable in the next release.

• Under the hood, FileAppenders use a java.io.FileOutputStream, which may be wrapped in a BufferedOutputStream. RandomAccessFileAppenders write to a ByteBuffer which is appended to the end of a RandomAccessFile when the buffer is full or flush() is called. There should not be much difference (apart from the performance) but FileAppenders have a longer track record and there may still be issues with the newer RandomAccessFileAppenders that haven't been discovered yet. Roll-over seems to work fine for both types of appender, but there may be corner cases that the team doesn't yet know about. (Log4J2 is under active development though and any issues will be addressed rapidly.)

• Both RandomAccessFileAppenders and FileAppenders can be configured with the immediateFlush attribute to flush every log event to disk. I recommend you switch off immediateFlush when using AsyncAppenders or AsyncLoggers so you can leverage the nice batching behaviour they provide: the async appenders and loggers can flush after logging multiple events and will also definitely flush once when the queue is empty again, which is very efficient and at the same time ensures that all log events are always persisted to disk.

Answer 2

Random access writes direct to the file with no buffering. This is slower but more reliable than buffered writing, in that buffered writing may lose a few unflushed writes in the event of a crash. You need to decide which is more important to you. It isn't clear from your list of three.

EDIT One would expect that a class called RandomAccessAppender would use random access, which implies no buffering, but apparently it doesn't!