>One common issue we have with that framework (and I assume we can have the >same with log4j2) is that all of our JVMs (we can have more than 50 JVMs on >the same server in production) roll their file at midnight. > >When this happens, the system became often not usable for a few seconds >because of the simultaneous zipping of all the rolled files that overload >the CPU (although zipping is done in a specific background thread).
ISTR that with the most recent versions of log4j, these threads are in a thread pool so that they are properly shutdown at the right time. I wonder if it's possible (or could be possible) to somehow inject a thread pool in to log4j for this rollover, so that for you use case you could inject a single thread executor, so only one thread is ever compressing at a time. Just a thought, anyway, Greg On 22 March 2017 at 13:47, Anthony Maire <maire.anth...@gmail.com> wrote: > Hi > > We are currently using another logging framework in production, but I'm > pushing to change it for log4j2. > > One common issue we have with that framework (and I assume we can have the > same with log4j2) is that all of our JVMs (we can have more than 50 JVMs on > the same server in production) roll their file at midnight. > > When this happens, the system became often not usable for a few seconds > because of the simultaneous zipping of all the rolled files that overload > the CPU (although zipping is done in a specific background thread). To > reduce this effect, we are combining a time based rolling policy with a > sized based policy to zip smaller files, but this is not enough to make the > system fully responsive at midnight. > > A pretty cool feature for us to avoid this issue is to have the possibility > when a rolling is triggered because of a time based policy to change file > immediately, but to wait for a random amount of time (within a configurable > limit) before starting the compression. This random delay should help a lot > to avoid contention on CPU cycles. > > Does log4j2 have something to solve this kind of issue ? If not, would you > accept a pull request for this (I will open a Jira if needed) ? > > Regards, > Anthony >