| 10 Apr 2026 |
 vcunat | Retrying isn't a real solution. Such tests should be either fixed or skipped. | 07:11:42 |
| vcunat | Otherwise we'll be dealing with the issue every rebuild again. | 07:12:06 |
| vcunat | * Otherwise we'll be dealing with the issue on every rebuild again. | 07:12:10 |
 Yureka (she/her) | The failing tests I'm seeing also look somewhat serious | 07:16:13 |
| Yureka (she/her) | Mismatch at indices:
[0, 1]: -0.07681634452326755 (ACTUAL), -0.19555901169952036 (DESIRED)
[1, 1]: 0.29433803718365575 (ACTUAL), 0.19727692219780213 (DESIRED)
Max absolute difference among violations: 0.11874267
Max relative difference among violations: 0.60719609
| 07:16:17 |
| Yureka (she/her) | this is not just some rounding error, but significantly outside of the range | 07:16:30 |
| Yureka (she/her) | I'll try to reproduce from master | 07:21:38 |
| Yureka (she/her) | to see if this has always been flaky/broken or just recently | 07:21:50 |
| Yureka (she/her) | Redacted or Malformed Event | 07:21:53 |
| Yureka (she/her) | but I feel like I've built scipy a lot of times already without these tests failing | 07:22:06 |
| Yureka (she/her) | worked on first try | 07:27:25 |
| Yureka (she/her) | let me try once more | 07:27:39 |
| Yureka (she/her) | but I really think it's something in this cycle | 07:27:46 |
 K900 | OK that looks sus | 07:35:44 |
| K900 | The ones that failed for me were like | 07:35:53 |
| K900 | Tiny errors | 07:35:55 |
| K900 | Is this on Asahi? | 07:35:59 |
| K900 | I wonder if you're hitting a different code path | 07:36:06 |
| Yureka (she/her) | 1x Asahi M1 Pro
1x Ampere One Debian | 07:36:49 |
| K900 | And it fails like that on both? | 07:37:14 |
| Yureka (she/her) | Yes | 07:37:16 |
| K900 | Oof | 07:37:22 |
| Yureka (she/her) | But both are non 4k pagesize | 07:37:22 |
| Yureka (she/her) | And it doesn't fail on the scipy from master | 07:37:35 |
| Yureka (she/her) | since it fails reliably for me, I can bisect it on my machine | 07:54:24 |
| Yureka (she/her) | I also think I managed to build it on a HoneyComb LX2K | 07:54:48 |
| vcunat | Maybe bisection won't be such a horrible experience. We now eval staging stdenvs every hour and for linux we do manage to build them all. | 08:01:58 |
| Yureka (she/her) | I think the actual problem is already present in openblas | 09:22:47 |
| Yureka (she/her) | didn't we have other openblas problems in this cycle too? | 09:24:27 |
| K900 | Something Darwin | 09:25:03 |
