Repository navigation
Editor play check flakes on linux-editor: a client's reconciled player 1.6 to 4.4 cm off the host's (limit 1 cm) #807
Description
Seen on linux-editor (ci.sh's editor play check, play_prediction_um against its 10 000 µm limit). The check fails now and then with nothing touching the editor or net code:
| run | branch | reconciled (prediction) | own | others |
|---|---|---|---|---|
| 39075192174 | main (push) | 33 005 µm | 0 | 34 |
| 39087186880 | main (push) | 44 214 µm | 0 | 50 |
| 39177636043 | feat/engine-pack92 (#806, an aephysics pin bump only) | 16 563 µm | 0 | 38 |
Main's later runs pass the same check. So a client's predicted player sometimes ends 1.6 to 4.4 cm off the host's after reconciliation, against the 1 cm the check holds (and that tests/test_players.ae holds). The other two numbers stay small: its own player at rest within 0, the others drawn within 34 to 50 µm.
Expected: the prediction within 1 cm on every run, or the check measured over a deterministic link if the error depends on the loopback's jitter and loss draw. The net flake approach (a loopback hub and a seeded sweep on a virtual clock, which turns intermittent failures into seeds that replay) should find the failing seeds.
Activity
nicolas-maman commentedon Oct 19, 2026
Again on PR #815 (c8e28e7, linux-editor, vulkan): "play: players off the host's -- reconciled 44214 um, own 0 um, others 50 um". Every other leg of that run was green; the job is rerun. The motion change in #815 is opt-in and off for the editor's figures, so this is the same flake.
nicolas-maman commentedon Oct 20, 2026
Again on PR #826 (motion and physics only, nothing in net or the editor), run 39352917287 at c5a67f7, linux-editor:
FAIL ae3d_editor (vulkan) (play: players off the host's -- reconciled 33516 um, own 0 um, others 41 um). The next run's linux-editor (39355383288, e3d87b2) passed.