I've been caught out with Nagles Algorithm before, it's one of the first thing's I disable now.
Been running in production with a persistence profile configured for about a month now (default timeout set - 180 seconds). I see the occasional error on the services, but this may possibly be narrowed down to kerberos timeouts due to inactivity from low client volume. I expect the client volume to ramp up and these errors will no longer exsist (tokens will be renewed regurally) over the next few months.