Would there be one registry or one per logger context? Gary -------- Original message -------- From: Remko Popma <[email protected]> Date: 07/04/2015 05:41 (GMT-08:00) To: Log4J Developers List <[email protected]> Subject: Re: Log4j2 RollingFileAppender deadlock issue
ThreadLocal is implemented as an internal Map in each Thread instance, so there is constant lookup time regardless of the number of Threads and the number of lookups. Contrast this with a lock, where performance will decrease exponentially with more concurrent threads. (See also https://plumbr.eu/blog/java/how-is-threadlocal-implemented ) Sent from my iPhone On 2015/07/04, at 20:40, Jess Holle <[email protected]> wrote: On 7/4/2015 2:51 AM, Gary Gregory wrote: On Thu, Jul 2, 2015 at 6:18 AM, Remko Popma <[email protected]> wrote: Yes, that could still work: We could use a ThreadLocal containing a custom class which holds the lastTimestamp, cachedDateString as well as a SimpleDateFormat instance. As Jess pointed out, we would also need a way to clear the ThreadLocal when the LoggerContext is stopped (to prevent memory leaks in web apps). This may be the third usage of ThreadLocals in log4j2 now, so it may be worth creating a reusable mechanism for this. One idea would be to have a ThreadLocal registry in the LoggerContext, where the LoggerContext is responsible for cleaning up all registered ThreadLocals in its stop() method. Thoughts? I'm wondering what the performance cost are of doing a ThreadLocal.get() vs. synchronized(this) on each call to format(). Personally I'd be less concerned with optimizing maximum logger throughput on any given thread than: Ensuring that not logging takes minimal time Minimizing potential thread contention Logging at maximum efficiency is a priority, but comes after these others. -- Jess Holle
