Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Ask HN: Ever worked with a service that can never be restarted?

news.ycombinator.com

Post reply on HN
Print view View original

31–40 of 201 posts

Re: Ask HN: Ever worked with a service that can never be restarted?

#31

Post by sudhirj »

gberger wrote: post #20
jomkr wrote: post #9

I'm not sure that 55 months of up-time indicates it's more or less likely to go down in the next month, but I'd guess more likely. Surely there are options, have you tried de-compiling the source from the binary? >but the physical config was accidentally overwritten and there are no backups Any old dev PCs lying around somewhere? It's worth reaching out to the old developers to see if they have a copy, in really old…

> I'm not sure that 55 months of up-time indicates it's more or less likely to go down in the next month, but I'd guess more likely. If 'going down' is an event governed by chance (e.g. power failure), then it does not matter if it has been up for 55 months or 55 minutes. https://en.wikipedia.org/wiki/Gambler%27s_fallacy

There’s also the Turkey fallacy - Turkeys that are smug about knowing the Gambler’s fallacy keep thinking their survival is independent of their age, until people eat them. My probability of dying goes up with each year as well, it’s not constant.

Think the probability of the machine shutting down in the next minute is independent of age, but it shutting off in the next 100 years is certain. There’s a line or curve that should indicate probability of shutting down before OP writes the replacement.

Re: Ask HN: Ever worked with a service that can never be restarted?

#32

Post by eru »

gberger wrote: post #20
jomkr wrote: post #9

I'm not sure that 55 months of up-time indicates it's more or less likely to go down in the next month, but I'd guess more likely. Surely there are options, have you tried de-compiling the source from the binary? >but the physical config was accidentally overwritten and there are no backups Any old dev PCs lying around somewhere? It's worth reaching out to the old developers to see if they have a copy, in really old…

> I'm not sure that 55 months of up-time indicates it's more or less likely to go down in the next month, but I'd guess more likely. If 'going down' is an event governed by chance (e.g. power failure), then it does not matter if it has been up for 55 months or 55 minutes. https://en.wikipedia.org/wiki/Gambler%27s_fallacy

It depends on what kind of chance we are talking about.

See eg https://en.wikipedia.org/wiki/Bathtub_curve for a common way to fail. Or https://en.wikipedia.org/wiki/Reliability_engineering in general.

You are right about some specific kinds of spontaneous failures. (Though not sure if external power failures are distributed that way.)

Re: Ask HN: Ever worked with a service that can never be restarted?

#33

Post by exdsq »

goatinaboat wrote: post #29

capture packet's and see what kind of traffic it receives most commonly The modern solution would be to capture the incoming packets as a training set then apply machine learning to create a model that can perfectly recreate the outgoing packets. It’s still an inexplicable black box of course, that is the nature of ML, but at least you can run it in the cloud now.

Is this seriously done? I just completed a small literature review for neural program synthesis which uses inputs as training data to generate a program using neural networks that can generate the right outputs but it's a huge search problem and I assume for anything non-trivial this is out of our reach? As an example, Deepmind achieved binary multiplication using this technique in 2018 and this was notable.

Re: Ask HN: Ever worked with a service that can never be restarted?

#34

Post by listenallyall »

gberger wrote: post #20
jomkr wrote: post #9

I'm not sure that 55 months of up-time indicates it's more or less likely to go down in the next month, but I'd guess more likely. Surely there are options, have you tried de-compiling the source from the binary? >but the physical config was accidentally overwritten and there are no backups Any old dev PCs lying around somewhere? It's worth reaching out to the old developers to see if they have a copy, in really old…

> I'm not sure that 55 months of up-time indicates it's more or less likely to go down in the next month, but I'd guess more likely. If 'going down' is an event governed by chance (e.g. power failure), then it does not matter if it has been up for 55 months or 55 minutes. https://en.wikipedia.org/wiki/Gambler%27s_fallacy

Gambler's Fallacy is applicable "when it has otherwise been established that the probability of such events does not depend on what has happened in the past" -- not the case here.

The fact that the server has been running uninterrupted for 55 months is a strong clue that the power to this system is robust, not subject to "random" power loss scenarios like one would expect at a typical residence.

Re: Ask HN: Ever worked with a service that can never be restarted?

#35

Post by goatinaboat »

exdsq wrote: post #33
goatinaboat wrote: post #29

capture packet's and see what kind of traffic it receives most commonly The modern solution would be to capture the incoming packets as a training set then apply machine learning to create a model that can perfectly recreate the outgoing packets. It’s still an inexplicable black box of course, that is the nature of ML, but at least you can run it in the cloud now.

Is this seriously done? I just completed a small literature review for neural program synthesis which uses inputs as training data to generate a program using neural networks that can generate the right outputs but it's a huge search problem and I assume for anything non-trivial this is out of our reach? As an example, Deepmind achieved binary multiplication using this technique in 2018 and this was notable.

I’ve never done it but it should be feasible. An IP packet is 1518x8 == 12144 bits which is not especially big for a one-hot style encoding these days. You would just need to be sure to capture representative traffic e.g. accounting for monthly jobs.

Re: Ask HN: Ever worked with a service that can never be restarted?

#37

Post by abbadadda »

I'm at the airport in Dublin. On a layover. I was enjoying my Oreo Fusion at BK and when I read, "The service has current up time of 55 Months" I burst out laughing hysterically.

(Update re the downvote: You can't tell me that uptime isn't a little bit funny. I almost did a spit take at this. As someone who sees and thinks an uptime of 150 days is way too much this number just shocked me, that's all)

Re: Ask HN: Ever worked with a service that can never be restarted?

#38

Post by rusk »

goatinaboat wrote: post #29

capture packet's and see what kind of traffic it receives most commonly The modern solution would be to capture the incoming packets as a training set then apply machine learning to create a model that can perfectly recreate the outgoing packets. It’s still an inexplicable black box of course, that is the nature of ML, but at least you can run it in the cloud now.

apply machine learning to create a model that can perfectly recreate the outgoing packets.

This reminds me of Team America [1] ... or maybe idiocracy [2] :-p

[1] https://www.youtube.com/watch?v=DIlG9aSMCpg

[2] https://www.youtube.com/watch?v=kIZ9YuPm_Ls

Re: Ask HN: Ever worked with a service that can never be restarted?

#39

Post by viraptor »

goatinaboat wrote: post #29

capture packet's and see what kind of traffic it receives most commonly The modern solution would be to capture the incoming packets as a training set then apply machine learning to create a model that can perfectly recreate the outgoing packets. It’s still an inexplicable black box of course, that is the nature of ML, but at least you can run it in the cloud now.

This is not a reasonable approach at this point. You'd need to know the whole state required for processing. That includes precise time, any local storage, any remote dependencies, randomness seeds, etc. Even then, we're not even close to full synthesis from examples. This is not a modern solution. This is made up, beyond toy examples.
Post reply on HN