After rolling back the PR I mentioned in the beginning of the topic we do not experience any issues. Even back then it was not "out of memory", but the process just ate one full core of CPU and never recovered. After all the CPUs are busy we cannot make any more queries/calls to JanusGraph.