Hacker Read top | best | new | newcomments | leaders | about | bookmarklet login

Another way to think of it might be that it can be relied on - until it can't. Be ready and wary of that happening, but until then you have what's probably a good mitigation of the problem.


sort by: page size:

You need to rely on something, not sure it's avoidable

Thus you can assume it won't be relied on or its flaws will be adequately mitigated.

More importantly the assumptions can be made about the tech. design and implementation and it's stability/availability across x. amount of time.

On the perfect part, I am not intellectually competent to answer that ...


That's a good point, I think I'd just heard that a long time ago and not really thought about it but with much larger features it should actually be more resilient.

Agreed.

Just think of the “maintenance” problem. Equipments do fail eventually


Interesting. I think it's a semantical battle between reliability (literally "can I rely on the system to be up ?") and resiliency ("if it takes a hit can I keep using it in degraded mode")

Once you decide that absolute reliability is unattainable, you can instead aim for sufficient redundancy. It’s not perfect, but with enough safe guards in place it may very well be good enough.

It might do well in these crises, but if it can't survive the times where everything works well, then it doesn't help.

Good point. Good example. I imagine it works well in my cases but I haven't come across cases like your example where it breaks down.

Yeah, that's roughly 99.99% availability, which sounds reasonable for most anything you want to depend on.

That's why fallback techniques like the ones in this article are a great idea (assuming that they are reliable).

Yeah, that's both good and bad — it's great that it's reliable but that's also the kind of thing I could see people getting complacent about because, hey, it hasn't failed before.

Exactly; a good system by definition should be robust to these kinds of things.

Two counter examples:

1. A cpu you rely on is a well understood system yet unexpected failures do happen (eg pentium bug, spectre exploits, etc).

2. A human you rely on might fail unexpectedly (tired, drunk, heart attack, going crazy, embracing terrorism, etc).

After testing reasonable number of things, if NNs perform more reliably than those other systems we rely on currently, it will be increasingly harder to justify using the other systems, especially when that means more people accidentally dying every year.


Something having a 30% chance of failing for a short period over 50+ years seems very reliable to me.

There are well known ways to build reliable systems out of unreliable ones through the use of slack and redundancy.

If you need something to live, you don't wait until you have a 48 hour supply and then rely on two day delivery, you keep a two week or more supply so that if delivery fails you have plenty of time to realize this and source it from somewhere else before it becomes an emergency.

And if it does become an emergency, there are emergency services for that. Go to a hospital and they will have or fly in just about anything. This is almost certainly far more expensive than not running everything down to the wire to begin with, but at least you don't die.


Does that matter if we're only going to use it in a situation like this? Some low reliability redundancy would still be nice

That's why you use it only when you can prove it won't fail. I'm ok with that, and said so at the start.

We have optimized for efficiency over resilience. It is great while it works but it makes the whole system more fragile.
next

Legal | privacy