Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
Starter 12
-
None
-
None
-
Mac OS, OpenJDK 1.8, Intel I9, 16gb RAM.
Description
I am testing performances of Apache Sling 12, configured with Oak Segment Store, started with the -Xmx4g parameter. I'm using a multi-threaded Python script (I tried with 30 threads down to 5 concurrent threads). The tool loads thousands of contents on new oak:Unstructured nodes using the standard Sling POST servlet. Before to start the import I stopped the Lucene bundle. Initially performances are good (about 200 new documents/sec). Soon, after the tool has ingested about 20k new nodes, the speed drops to a few nodes per second. I analyzed the JVM and noticed that the memory allocated for some threads grows by several GBs without ever going down (see attached images). If I stop the import tool, the situation does not change. The only way to solve this problem is to turn off the Sling instance and restart it. Is it normal for threads and in particular for the threads that implement the POST servlet, to acquire resources that are never released? I am available to reproduce this issue during a Teams / Google Meet.