If your monitoring can't handle your what system is doing, aggregate locally until it can. Having systematically incorrect stats (because loss is correlated with load) is worse than having none.