Major Description
We have recently upgraded from JanusGraph 0.6.2 (TinkerPop 3.5.3) to JanusGraph 0.6.3 (TinkerPop 3.5.5) and have observed an increase in all latency metrics reported by Gremlin Server.
Using a profiler we have seen that more time is spent in TraversalHelper.applyTraversalRecursively calling itself recursively multiple times.
Then we've narrowed down the issue to this change and resolved the latency problem by reverting the change in DefaultTraversal.java.
TLDR; as far as I understand, it looks like, in our case at least, this other round of recursion does have a significant performance impact. As I'm not too familiar with that part of TinkerPop code, it's unclear to me if that's because of our usage of JanusGraph/TinkerPop (feel free to advise). But in any case, I thought that was worth reporting in case others hit the same issue.
Happy to share/investigate more if needed.