Should probably expand on this a bit. In 2014 we were involved in a project with a reasonably well-known power grid operator. They had been deploying expensive equipment to monitor the status of various devices at remote substations. They had soft requirements for uptime, which means they set a budget and got the most reliable equipment they could afford. The monitoring system was expected to work somewhere between 4 and 5 nines, but if it didn't no one was going to get fired. The equipment they had been deploying was more or less expensive overkill, offering 5 9's for about $50k per box. Multiply that by about 1600 and you get the idea for the budget. Someone in the organization hit on the idea of using redundant consumer-grade equipment and initially spec'd out some rack-mount x86 systems with beefy redundant power supplies and SSD storage. But then they heard about RasPi's and wanted to give them a try.
Should you use a Raspberry Pi for that? The best way to answer that question was to try it out. The first 20 units were built out with a beefy, though small UPS and a decent enclosure. They were deployed alongside existing systems to see if they a) worked, b) gave the same data as the expensive system and c) were reliable.
The answer was... some individual units did not hit the 4 9's reliability target. The ones that did, seemed to continue to be reliable up to at least 5 9's. But it was hard to determine which would fail without putting them in the field, waiting six months and seeing which ones hung. But the price was cheap enough that putting two in the same corner of the wiring cabinet and adding a hardware watchdog was quite affordable.
We deployed just over 1000 in this redundant configuration and it worked fine.
By this time we collected enough data to chart a MTBF histogram, essentially a chart of how many machines lasted how many days without needing a reboot. We wound up using a lot of Original RaspberryPi B+'s in 2015, which was after the RasPi 2 came out. (Purchasing at this client took a LONG time.) I often wonder if our supplier sent us boards that had been returned from other customers.
I had good experiences with BeagleBoard's in the late 2000's and this was just as the Black was coming out. We deployed another 600 with BeagleBone Blacks and got MUCH better reliability numbers. Is the BBB an intrinsically better product? I dunno. Did I just get a batch of bad RasPi's from my supplier? I dunno. Should you ever run an embedded system without a hardware watchdog? Probably not, no matter how reliable you think your system is.
But... the real problem with the RasPis was support. How do you return a RasPi? You don't. You throw it away and get a new one. Yeah. That doesn't work for a lot of people. How do you get the firmware for a RasPi in 2015? You don't. Can I get the gerbers so I can turn a few custom boards? No. I talked with RPT several times about this and their response was "we're not an ODM."
And that's PERFECTLY FINE. I never said I though RasPi's were "bad" -- I may have said "there are applications for which RasPis are not a great fit." If you need the firmware, consistent reliability or the ability to turn custom boards, you absolutely don't want to buy a raspberry pi in 2015.
Also... a lot of people are responding with "Except for the several times the system rebooted, I didn't have to reboot my RasPi," which I'm not sure I understand.