Danfra I didn't say it was less. I said the VRAM math works out same because I owned the boxes before the fire. Migration cost was shipping, not new hardware. Cloud A100s are 3-4x what I pay for space and power. Do the math on 24 months.
Garykwh this is the spreadsheet I wanted. Owned hardware + colo vs cloud rental, breakeven around month 14-18 depending on power deal. I have this for Vancouver hydro rates vs Seattle. Same story.
Member OP
Back to traceroutes. I re-ran mine this morning.
[code]Vultr Frankfurt: 18.7ms now, route via 198.244.x.x still but different intermediate
Contabo Munich: 14.3ms, unchanged
Hetzner Falkenstein: 13.0ms, unchanged
OVHcloud Gravelines: 12.1ms, one hop longer
Vultr improved 4ms. Still not back to pre-fire path but they moved something.
I test restores, not routes. Three years ago my OVHcloud backup was in Strasbourg. Same fire region. Now I keep one copy with Hetzner Helsinki, one with Contabo Singapore. Restore test last month: Helsinki 4 hours, Singapore 11 hours. Both worked. That is my concrete number.
Member
Readerkrakw 11 hours from Singapore? What size dataset?
Tomchan 4.3 TB compressed. My pipe is the bottleneck not theirs.
Same substation story was Dallas. KnownHost and another provider I won't name. Both "diverse" racks, same ERCOT event 2021. Not fire but same lesson. Nobody publishes substation data because nobody asks in sales chat.
Joe27 how did you find out same substation? FOIA? Utility maps? Or after it failed?
AblazeHades after. Outage postmortem from one provider mentioned substation ID. I recognized it from the other's status page. They edited it out later but I had screenshot.
This is why I tell junior SREs: geographic diversity means different utility, different flood plain, different seismic zone. Same metro fails two of three. I have runbooks that specify minimum 300km for stateful, 100km for stateless. Most providers "zones" are 30km.
