before looking at cluster node performance, did you check if the custom service configuration is a sane one?
E.g. not instrumenting something like a massively called method or a method that leads to very long running purepaths (e.g. a batch job that would run for hours and create 100s of thousands PP nodes maybe?)
request naming rules will impact the “grouping” of the requests not necessarily the amount of PurePaths created. You will get fewer buckets but these buckets will hold more PurePaths.
Can you maybe show a concrete example what you have instrumented and what the resulting PurePaths/Requests are? That would make it easier to help