ship-model · diff
git:20260922.754ed13 to git:20260922.a4bd47b
2 added, 2 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
---
name: ship-model
description: >-
- Release a retrained model into serving. Trigger on "выкати модель",
- "задеплой модель", "ship this model", "publish the new weights", or after a
+ Release a retrained model into serving. Trigger on "ship this model",
+ "publish the model", "publish the new weights", or after a
model change has passed the eval gate. Covers snapshot, registry, artifact
publish, verification against live, and the rollback path.
---
# Ship a model
Serving loads artifacts from storage, not from the training repo's working
tree. Most model releases that go wrong go wrong in that gap.
## Steps
1. **Refuse to start without a passing gate.** The `model-eval` verdict, with
its pooled and per-season numbers, is the entry condition. No verdict means
no release.
2. **Snapshot the model currently in production first.** The rollback plan is a
command you have already tested, not a paragraph of intent.
3. **Bump the version in the registry** and record what changed: features
added or removed, the eval verdict, the date. The registry is how anyone
answers "what is actually live right now" a month from now.
4. **Publish artifacts to the exact paths serving reads.** Same bucket, same
key layout, same bundled data files. A model published next to the right
path is a model that is not live.
5. **Respect the deploy order** from the feature change: if the feature list
moved, serving and the artifact have to land in the order the PR stated.
6. **Flush the caches between you and the user.** Prediction output is usually
cached in more than one layer, and a correct new model behind a warm cache
looks exactly like a release that did nothing.
7. **Verify against live, not against the test suite.** Request a real
prediction from the deployed service and compare it with what the new model
produces locally. Equal means shipped. Different means the old artifact is
still being served.
8. **Say which model is live** in the release note, with its version and the
one-line reason it replaced the previous one.