My work machines are "mission critical", configured to sustain three or four "nines" of uptime, built with as much "server grade" reliability and redundancy as possible. So these machines use ASUS boards and Titanium PSUs and enterprise-class SSDs and quality UPS backups and a robust isolation transformer, with ECC memories and self-powered data caches wherever they're supported. I consider overclocks "stable" on these machines when they can run brutal stress/torture tests for a full week (168 hours), uninterrupted, unsupervised, in a nasty "worst case" 30C to 35C ambient. (Although I do check their status a few times each day, just to confirm they're not crashed or catching on fire, to not waste more time on already failed test burns, but I don't really have to do so. Data corruption is absolutely intolerable, data integrity is highest priority. Downtime is somewhat intolerable, lost uptime equals forced downtime which I'd rather spend doing useful work than spending on troubleshooting/tinkering a failed platform. Yes, it means the overclocks are more marginal/modest and less reckless/spectacular.)

My overclocked gaming machines only need to stress for a few hours, I usually just run them overnight. The logic is that I'm only going to actually demand full performance across the entire duration of a long gaming session (a few hours at a time, at most). A BSoD/crash/lock/reboot or other fault which appears only rarely and only becomes evident after days of relentless punishment is of little concern, especially since whatever data's on the machine isn't "critical" and it can all be restored/rebooted to working order quick enough. So I punch up the overclocks towards the maximums the machines can sustain, turning them down a notch only if crashes/etc become annoying. An important part of this "maximum overclock" approach is rigid adherence to a solid backup strategy.

So it's a personal judgement call. From what I've read online, most gamers/overclockers seem to consider one hour or a few hours sufficient, even ten to twenty minutes is enough to weed out the overwhelming majority of unstable overclocks. It's also not as simple as measuring a definite "pass/fail" threshold, each part has a blurry zone where a decision needs to be made about how far up the curve of performance vs temps/volts/stability/longevity you want to go, the more you shift the balance to parameter(s) on one side the more sacrifice to parameter(s) on the opposite side. Again, it's a personal judgement call.