| 21 Aug 2025 |
 K900 | (and probably should be done outside official infra, at least initially) | 15:46:06 |
 vcunat | Trust is needed. People who have physical access to the HW can more-or-less put anything into the cache. | 15:53:05 |
| vcunat | (though not trivially at least) | 15:53:34 |
 Tristan Ross | If the infra team doesn't have physical access but the hardware team does, is that acceptable for trust? | 15:53:55 |
| K900 | Would not be great | 15:54:38 |
| K900 | But also, again, I want to first see proof that there's enough hardware | 15:54:38 |
| K900 | And that there's enough demand to justify the Hydra costs | 15:54:43 |
| Tristan Ross | What sort of costs does it make on Hydra? Is it just eval and storage? | 15:55:56 |
| K900 | Queue runner is slow | 15:56:35 |
 emily | cache for the entire jobset on a new platform is substantial (since we pay the whole cost every staging cycle) | 15:56:40 |
| K900 | Queue runner slowness scales with number of jobs | 15:56:44 |
| emily | and Hydra scheduling is a bottleneck | 15:56:53 |
| K900 | Storage is also still questionable | 15:56:59 |
| K900 | Eval is largely not an issue | 15:57:06 |
| emily | I think it would make sense to prove this externally first | 15:57:12 |
| emily | with a separate Hydra | 15:57:23 |
| emily | and see if the hardware can even keep up | 15:57:33 |
