Have you checked for memory growth over time?

Try running the main bulk of your code, the checks, in a standalone script that doesn't use preforking, see whether that changes things.

If you can just loop calling the checks as fast as your network admin will let you get away with, on a test network if such is available, or after hours, weekends, whatever makes sense in your environment. Perhaps only using a single machine as the target of your tests, maybe your own. The idea being to run the components as fast as you can to artificially excaserbate the problem. If you can get the problem to occur in a reasonably short period of time, then it becomes much easier to exclude the various subcomponents one at a time and try and isolate the fault that way.