I guess if they run a 0.001% risk of doing something disastrous then it doesn't cut it. If 0.001% of purchases fail or orders get lost, or they have a few minutes of complete downtime, it might not be worth adding another 9.

It says something interesting about how think about "high reliability." What it really means is "must never, ever fail" which kind of makes "% reliability" a sort of strange euphemism, I think.

High reliability means that it is much more likely that something else fails first. We could all be wiped out by a gamma ray burst for example, it makes little sense to make something last longer in expectation than the expected time between two major extinction events.

What? $1B over a day vs $1B over a year is the same with regard to 99.999%.

If your timeline is only one day, 99.999% reliability is less than a second of downtime.

I think his point is if you're making $1B a day vs $1B in a year downtime suddenly becomes much more costly for you

.001 of a years worth of minutes is 525 minutes (.001 * 525600) or about 8 hours and 40 minutes. If all of that time just so happened to occur on the $1B day then that's 356 million. Of course this is back of the napkin. That $1B is not evenly distributed throughout the day and especially in high availability systems the .001 downtime is many small outages over the entire infrastructure and not the entire apparatus crashing all at once. Though because load can be a factor in downtime, they are not completely without correlation.

On mobile but didn't you leave off the percentage sign? (which is another two zeros after the dot)

Of course this is taking op very literally

EDIT: I checked: google "0.001% * 1 year" and it tells you 5.259 minutes, so you were off due to dropping the % sign. That changes things dramatically: beyond the 100x difference, there is very high likelihood that someone would check back 2 minutes later (especially if they are waiting all year), so 5 minutes of downtime (spread over a whole day) likely objectively results in much lower lost sales than your calculation implied! For future envelope calculations don't forget that a % sign is exactly moving the decimal two more places! 55.2% is exactly .552 and so forth!

No, my point is that 99.999% uptime has to be scoped to the same timeline as your window that you're measuring. If you're focusing on availability for a single day, then achieving 99.999% availability for that day means less than a second of downtime.

I don't even monitor for sub-second downtime... A sub 500ms glitch would very likely go unnoticed on most of my production platforms. @$1B/5mins, that could represent ~$1.5million in (potentially) lost sales...

That's just a $10K loss.

It is not that hard. I work on a similar system, with a demand of maybe 30 minutes(take a guess, there are not may of us) over the entire year.

What I do is, have an active-active system in a data center, and with another similar system in another DC.

Having said that, your average online stock brokers do the same thing.

30 minutes once a year? That sounds like my sex life