I've found the Pis to be surprisingly resilient, heat-wise. I built a handful of inline monitoring servers running suricata, pulling yml configs from puppet, and forming an ad-hoc network with one another, with devices in a reasonably harsh manufacturing environment for six months' testing (near welding robots, and in network cabinets that average 180F during "non-load" hours) and didn't see a single system failure, crash, or reboot during that time. Literally the only thing I had problems with was maintaining consistency for the ad-hoc network, and that was largely owed to the amount of interference on the manufacturing floor, combined with greater-than-suggested distances between the devices.