Turns out "it's not a bug, it's a feature". In a nutshell, JMX causes quite excessive allocation, (and it's a known problem and it's especially visible with VisualVM (with which we end up with constant 10k/s objects allocations). The recommendations are: decrease pooling interval (doesn't help that much with VisualVM) or use other tools (YouTrack operates in native memory and doesn't impact heap in this case)

In the attached screenshot:

1. start jconsole
2. start visualvm
3. stop visualvm
4. start tigase-monitor (which triggered GC)
5. start visualvm
6. stop visualvm

In conclusion - using jconsole or tigase-monitor impacts (memory) metrics the least and this increased memory usage has to be taken into consideration when monitoring using JMX. (Alternatives like prometheous #issue #1034 might be better here and could possibly be considered in the context of Cangaroo and it's monitoring (#kangaroo-29) ; there is a JavaFlightRecorder, which is supposedly low-impact monitoring solution, but the tooling around it is not great)