Are servers based on consumer grade hardware reliable enough for production?

Comments

• MikeA Patron Provider, Veteran

  5:00AM edited 5:04AM

  doubt much of a difference. the main problems usually stem from prosumer type NVMe drives for me. but I've also had expensive hardware, or things like Supermicro boards, be very problematic. it probably doesn't matter what you buy, it just comes down to luck. you pay more to have a easy replacement, so enterprise don't complain about problems because they get them replaced easily (well, maybe not now days, unless you are OpenAI or Anthropic or Elon).

  Also, having proper cooling is a huge thing. If components are cooking, shit hits the fan.

  Thanked by 1iriska
• Advin Member, Host Rep

  5:03AM edited 5:08AM

  I've seen some wild issues on EPYC but it has overall been more stable/predictable than Ryzen. We've had some really weird issues on Ryzen with virtualization in the past and often times we wouldn't have any idea as to what was going on (e.g., random freezing, reboots without explanation, etc). At the same time, we've had issues with EPYC servers in the past.

  We do have a bunch of Ryzen servers in our fleet that haven't had issues since day 1, and we are comfortable running some production services on Ryzen processors. It's somewhat of a rarity for there to be issues as the platform has matured, but when there are, they are generally just harder to diagnose/usually some much more obscure issue.

  I feel like hardware has gone downhill in terms of stability with how these platforms are being pushed to their limits (e.g., DDR5 at extreme speeds, high core processors, insane amounts of PCIe, etc).

  Thanked by 1oloke
• kukara4 Member, Patron Provider

  5:11AM

  I would say it’s more about architecture. Anything with a single copy is not production, even on enterprise servers.

  Thanked by 1MikeA
• waqid Member

  6:00AM

  Epyc has more testing in fabs hence more stable than ryzen
• xmrlist Member

  7:07AM

  Agree with kukara4 that redundancy matters more than the badge, but a few concrete differences are worth checking before you put something important on a consumer box:

  • ECC: lots of Ryzen boards accept unbuffered ECC but don't always report corrected errors. If the host advertises ECC, check dmesg | grep -i edac (or edac-util -v) to see if it's actually active.
  • NVMe: consumer drives usually have no power-loss protection and lower TBW. Fine for web/app servers, but for databases doing lots of fsync, enterprise drives are noticeably faster and safer. smartctl -a /dev/nvme0 shows wear ("Percentage Used"), so you can see what you were given.
  • Out-of-band management: most consumer boards have no IPMI/BMC, so a hard hang means waiting for the provider's KVM or remote hands. Worth asking what their response time is.
  • Burn-in: run memtest and a stress test on delivery; most hardware faults show up in the first days.

  If the app can survive one box dying (backups plus a second node or a quick restore), consumer hardware at a lower price is usually a good trade.