## TL;DR
Redeploying the model should never stop scoring. Serve blue-green: the old version keeps serving while the new one warms up, traffic flips only after the new version passes a health gate, and the old version stays warm as instant rollback. Signals keep flowing through the whole window with no agent-side changes.

```text
fraud agent timed out during model redeploy and missed fraud signals
```

## Use this when
- Scoring gaps line up with deploy timestamps.
- The agent times out waiting for the model endpoint during releases.
- A bad deploy takes down scoring until someone rolls back by hand.

## Not for this skill when
- The model is slow all the time. That is a capacity problem, not a deploy problem.
- Timeouts come from the network between agent and model. Check connectivity.

## Steps
1. Serve two versions during deploy: old keeps full traffic while new loads and warms its caches. Verify: latency and error metrics stay flat through a staged deploy.
2. Gate the flip on health: the new version must pass a canary scoring set before it takes traffic. Verify: a deliberately broken candidate fails the gate and never serves.
3. Keep the old version warm for fast rollback: one config change restores it. Verify: rollback completes inside the scoring SLA.
4. Make the agent deploy-aware: it retries against the stable endpoint alias, never a version-pinned host. Verify: agent calls succeed across a deploy with no code change on the agent side.

## Variant phrasings
### scoring downtime during model deploy
Availability phrasing.
### model redeploy dropped fraud signals
Signal-loss phrasing.
### agent timeout on model version switch
Timeout phrasing.

Compatibility: Works for hosted endpoints and self-served models. The blue-green pattern is the same whether the flip is DNS, config, or a gateway rule.

## Why it happens
Deploys were built as stop-and-replace, so scoring has a gap by design. The agent has no fallback and no idea a deploy is happening, so it times out and the signals in flight are lost.

## Edge cases / pitfalls
- Warmup needs realistic traffic shape, not just a health ping, or the first real burst is the warmup.
- Two versions cost double during the window. Budget the overlap.
- Feature stores must serve both versions or the new model reads stale features.
- Log which model version scored each transaction, or post-incident debugging gets impossible.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_PaX1ZTGjwdOb-SlnRwUy6A