3 ms·

I’ve been renting compute by the hour on Vast.ai and often wonder about the servers I use. Is it reasonable to assume that any such server with, say, a 90+% rel

I’ve been renting compute by the hour on Vast.ai and often wonder about the servers I use. Is it reasonable to assume that any such server with, say, a 90+% reliability rating is in a data center, rather than someone’s basement?

j45 2y ago
Not really, the equipment is the main point of failure. Then maybe the connection and power, and environment (cooling) The last 2 are really well solved. It's crazy what's possible with USFF machines as mini-blade servers. Low electricity and really serviceable redundancy and failover.
- everforward 2y ago
  This. 90% gives surprisingly large error budgets; that’s about 72 hours of downtime a month or 36.5 days a year. I would guess most home labs are at or near 99% availability on hardware. Scale is part of it too. I could pack my homelab into the car and take it to a friends house. They have the space, power service, and internet to accommodate me. A full data center doesn’t have that option, at least not for free-ish.
  - j45 2y ago
    Haha, thanks. When you can buy 4 USFFs for the price of 1 new one, and run 2 have have a test one, and one for spares I'm not really sure what everyone is concerned about. Enterprise, engineering, industrial grade equipment is the same high quality level, be it a server or a desktop meant for extreme environments. Too many people are only purchasers of consumer grade equipment and would be surprised at what a huge difference the corporate/engineering/industry spec stuff is like.
omneity 2y ago
vast.ai have two kinds of servers, what they call "datacenter" and regular ones. Datacenter servers seem to have better internet and a contractual agreement with vast.ai, probably the distinction you're looking for.
nfriedly 2y ago
Eh, the server in my basement hits one 9 of reliability. Maybe even two!

3 ms·

I’ve been renting compute by the hour on Vast.ai and often wonder about the servers I use. Is it reasonable to assume that any such server with, say, a 90+% rel

by munchler 2y ago

j45 2y ago
Not really, the equipment is the main point of failure. Then maybe the connection and power, and environment (cooling) The last 2 are really well solved. It's crazy what's possible with USFF machines as mini-blade servers. Low electricity and really serviceable redundancy and failover.
- everforward 2y ago
  This. 90% gives surprisingly large error budgets; that’s about 72 hours of downtime a month or 36.5 days a year. I would guess most home labs are at or near 99% availability on hardware. Scale is part of it too. I could pack my homelab into the car and take it to a friends house. They have the space, power service, and internet to accommodate me. A full data center doesn’t have that option, at least not for free-ish.
  - j45 2y ago
    Haha, thanks. When you can buy 4 USFFs for the price of 1 new one, and run 2 have have a test one, and one for spares I'm not really sure what everyone is concerned about. Enterprise, engineering, industrial grade equipment is the same high quality level, be it a server or a desktop meant for extreme environments. Too many people are only purchasers of consumer grade equipment and would be surprised at what a huge difference the corporate/engineering/industry spec stuff is like.
omneity 2y ago
vast.ai have two kinds of servers, what they call "datacenter" and regular ones. Datacenter servers seem to have better internet and a contractual agreement with vast.ai, probably the distinction you're looking for.
nfriedly 2y ago
Eh, the server in my basement hits one 9 of reliability. Maybe even two!