Dmitry Konstantinov commented on LOG4J2-2269:

Please find in attach a code to reproduce the problem - 
[^test_to_reproduce.zip]. It is not pure Unit test, it is quite tricky to 
reproduce the problem in format of a usual unit test because the observable 
effects are side ones.

If we run the test with -Dlog4j2.enableThreadlocals=true (default) we got a 
result like: 
Used heap size: 101 Mb, max heap size: 245 Mb 
at the end of execution.
For  -Dlog4j2.enableThreadlocals=false the same test gives:
Used heap size: 3 Mb, max heap size: 245 Mb
Heap dumps for both cases are also attached to the ticket:  
[^heap_dump_threadLocals_false.hprof.zip] ,  

In heap dump  [^heap_dump_threadLocals_true.hprof.zip]  we see that the 
application level heavy objects (TaskContext) are kept in memory because of 
log4j2 references in thread locals:

> Memory leaking of replacement parameters in case of enableThreadlocals=true
> ---------------------------------------------------------------------------
>                 Key: LOG4J2-2269
>                 URL: https://issues.apache.org/jira/browse/LOG4J2-2269
>             Project: Log4j 2
>          Issue Type: Bug
>          Components: Core
>            Reporter: Dmitry Konstantinov
>            Assignee: Remko Popma
>            Priority: Major
>         Attachments: heap_dump_references.png, 
> heap_dump_threadLocals_false.hprof.zip, 
> heap_dump_threadLocals_true.hprof.zip, test_to_reproduce.zip
> Use case:
> 1) I have a thread pool with quite large number of threads (around 100)
> 2) Threads from the pool sometimes process some heavy objects. Threads are 
> not actively writing into logs (so, a thread may process a task and does not 
> log any message).
> 3) As a part of processing for some objects in some non-frequently used 
> branch a thread logged a message with heavy replacement parameters 
> (parameterized log message is used).
> In this case such heavy replacement parameters are not garbage collected 
> until another message will not written within the same thread because 
> thread-local ReusableParameterizedMessage and MutableLogEvent objects keep 
> references to the replacement parameters.
> Is it possible to cleanup such references when a message is released (for 
> example as a part of 
> org.apache.logging.log4j.message.ReusableMessageFactory#release and 
> org.apache.logging.log4j.core.impl.ReusableLogEventFactory#release) to avoid 
> such kind of leaks?

This message was sent by Atlassian JIRA

Reply via email to