3 ms·
While the pricing is compelling, most customers want at least 98℅ of the reliability that industry giants offer. Maybe the community can comment on whether pack
by mankash666 10y ago
While the pricing is compelling, most customers want at least 98℅ of the reliability that industry giants offer. Maybe the community can comment on whether packet.Net is reliable enough to trust your business with?
- zsmith928 10y agoFYI I'm from Packet. Agreed reliability is important (I'll let others comment on our own record) but I'd say there is significant workload via spot market, batch processing, etc that doesn't need uptime and is interested in cost. Also, I've seen most users want 99.9+% uptime. 2% downtime is 14 hours in a short month. People would probably get angry about that...
- shadeslayer 10y agoHey Any chance you guys are going to offer something that is on the lower end of the spectrum for ARMv8 in terms of pricing? For eg. Scaleway has ARMv7 servers for 3 EUR/month (https://www.scaleway.com/pricing/ https://www.scaleway.com/pricing/)
- zsmith928 10y agoYes, we are considering a Type 0 ARM server, but we are still working through the proper design to enable full Armv8 (vs v7).
- shadeslayer 10y agoSweet :D I'll keep an eye on the Packet website. I did try out the ARMv8 server, a bit disappointed that the hardware did not support backwards compatibility with ARMv7 though :(
- dylz 10y agoLooking for something similarly priced to scaleway in US too. Have a ton of services with them, but can't find an alternative at lower latency.
- jzelinskie 10y agoAt CoreOS, we use Packet to host the Quay.io build system[0]. Getting bare-metal machines from Packet that we can run our own virtualization on has saved us a significant portion of our expenses compared to our previous system which dynamical spun up EC2 instances. This is definitely not the traditional deployment, but if you're using things like dedicated servers on AWS, you should be evaluating Packet. In my own personal ops experience with Packet, I've actually seen less failures than Amazon; it feels like more than once a week Amazon has degraded service on the EC2 API in us-east-1 that used to cause us downtime. Not to mention, Amazon waits like 4 hours before they admit anything ever went wrong (green with a tiny yellow badge). However, I'd take this with a grain of salt, as we've replaced the Amazon control plane with a Kubernetes control plane and only interact with Packet API when we tear-down/spin-up a new cluster. [0]: https://blog.quay.io/faster-container-builds https://blog.quay.io/faster-container-builds
- quitlahok 10y agowe use packet.net for all of our services at andyet.com and i can only remember one time that we had a problem that was actually because of them and not user error. as i recall they had an issue with some networking hardware, but had it repaired in about an hour. i've seen a few other times where they've mentioned issues in their status page, but they're typically only for low inventory, and even those are usually resolved pretty quickly. overall we're super happy with packet and i recommend them to anyone looking for hosting - especially those that have a need for bare metal servers.