I'm aware of several Java service backends that just disable the GC entirely and take a once-a-day crash as a cost of doing business. As long as you have enough nodes that the crashes aren't correlated, your overall service can maintain decent uptime