Reference
Service levels
What we promise for each plan, the credits when we miss, and how maintenance works.
By plan
| Plan | Promise |
|---|---|
| Anvil | 99.7 % of minutes in a calendar month, per node, with credits |
| Strike and Hour bank | Jobs ended by our hardware are not billed, and bank hours are returned |
| Ember | None beyond the 30 second notice |
What counts as down
For an Anvil, a minute is down when a GPU on it cannot run your job for a reason on our side: a hardware fault, a host fault, a network fault or maintenance we did not announce.
Credits
| Anvil availability in the month | Credit on next invoice |
|---|---|
| Under 99.7 % | 10 % of that node's month |
| Under 99.0 % | 25 % |
| Under 97.0 % | 50 % |
Credits are applied by themselves. They are the only remedy.
Maintenance
Announced at least 7 days ahead in the incident log, at most once a month, at most 3 hours, starting between 02:00 and 05:00 UTC. Announced windows do not count as down.
Faulty GPUs
A GPU that reports uncorrectable memory errors, falls off the bus or fails a health check leaves scheduling within 60 seconds. Jobs on it end with result hardware. Anvil teams get a spare GPU from the shared pool when one is free; if none is, the minutes count as down.
Support
| Plan | First response |
|---|---|
| Anvil, hardware | Phone, any hour, 30 minutes |
| Anvil, other | Email, 2 business hours |
| Strike, Hour bank, Ember | Email, 1 business day |
Business hours are 08:00 to 18:00 UTC, Monday to Friday.