steveipw
Member
OP
- Joined:
- Jun 2024
- Posts:
- 163
- From:
- Denver, US
I run a small ZFS pool for video renders at two of your locations. Before signing, I read the SLA three times because I've been burned before.
The document promises <15 ms between "designated metro pairs." Fine. I check: my traffic between your East-1 and Central-2 regions goes East-1 → PeerA → TransitB → your own backbone → Central-2. Traceroute shows 34 ms. MTR shows 4% loss during peak.
I open a ticket. Your NOC replies: "The SLA only covers direct backbone paths. Your flow is routed via cost-optimized peering."
The route you actually use. The one you actually route me on. Excluded by definition.
I asked for clarification. Your rep admitted: "That wording is legacy from before we restructured the network." No fight, no escalation. They offered a 15% service credit without me asking.
For that money you get 34 ms and a shrug. I'm not even angry, just documenting for the next person who reads before signing.
Current setup if anyone cares:
- Primary: 8x14 TB Toshiba MG08, RAID-Z2, ~70 TB usable
- Replica to your Central-2 every 6 hours via rsync+ZFS send
- For that money you get a used Dell R740xd at auction: €2200, 128 GB RAM
The credit covers roughly one month of the 10 Gbps link. I'll take it.
zfs send | zfs receive. repeat.
JokoNord
Member
- Joined:
- Jun 2024
- Posts:
- 190
- From:
- Jakarta, ID
RTFM
Yaml
Sla:
Latency:
Guaranteed_ms: 15
Applies_to: "direct_backbone_only"
Actual_route: "see: appendix_z, section_4"
- the fine print giveth
- the fine print taketh away
- read appendix z
Your 15% credit is them admitting the yaml is broken
traffic worse than my packet loss
hankels
Member
52 VPS and counting
- Joined:
- Jun 2024
- Posts:
- 301
- From:
- Phoenix, US
I run 52 boxes across 7 providers and I track latency SLA in a spreadsheet
Contabo: no SLA at all, honest about it
Hostinger: 20 ms, actually hits 12
Your provider: 15 ms, gives 34, calls it "legacy"
My rule: $/GB/RAM + latency penalty if they lie. You're getting what, 0.4 penalty? Could be worse
seedbox, NAS, tape, and three offsite
garykwh
Member
Inference on a Budget
- Joined:
- Jun 2024
- Posts:
- 173
- From:
- San Jose, US
4% loss and they still call it "legacy"? I'd be on the phone
CUDA cores are my love language
steveipw
Member
OP
- Joined:
- Jun 2024
- Posts:
- 163
- From:
- Denver, US
Update: they offered 15% service credit. I asked how that's calculated. "Against the compute portion of your monthly invoice, excluding bandwidth, storage, and support fees."
So on a $340 bill, that's maybe $12. For a route that runs 34 ms with 4% loss against a 15 ms promise.
I told them to keep it and fix the doc instead. They closed the ticket "resolved - customer declined credit."
zfs send | zfs receive. repeat.
garykwh
Member
Inference on a Budget
- Joined:
- Jun 2024
- Posts:
- 173
- From:
- San Jose, US
That's insulting. 15% of compute only? My GPU inference jobs live or die on that latency. You should name them so the rest of us know who to avoid.
CUDA cores are my love language
hankels
Member
52 VPS and counting
- Joined:
- Jun 2024
- Posts:
- 301
- From:
- Phoenix, US
He won't, and I get it. They're big enough to make life hard. But I will say: if your SLA has an "Appendix Z," run.
My Contabo boxes in Seattle to Phoenix: 42 ms, no SLA, no lie. My Hostinger box in Phoenix to their Paris DC: 138 ms, and they told me upfront it would be. Honesty is a feature.
seedbox, NAS, tape, and three offsite
