▱ModelLaunch/
ML infrastructure workbenchGitHub ↗
Local Kubernetes · model delivery

Ship a model. Prove it holds.

Replay real releases recorded on a local kind cluster: a gated rollout, a rejected candidate, and an automatic rollback measured from breach to recovery.

Replaying committed release records
Select a record
-
Active model
-
Stable traffic destination
p95 latency
- ms
Prometheus 30 s window
Request errors
-
5xx / all, 30 s window
Ready replicas
-
Candidate deployment

Where traffic goes

Stable route
Stable service/predict
selector: none
Previous deploymentServing
-
-
Candidate deploymentPending
-
-

Candidate checks run behind the internal preview Service. The previous deployment stays ready through the verification window; promotion and rollback are a selector patch on the stable Service.

One release workflow

0 / 6 STEPS

Service health through the release

Prometheus samples (watch)   preview   recovery   ┄ gate

Release activity

RECORDED EVENTS

    Release contract

    Fixed before the run
    Quality metric / validation set
    -
    Candidate validation result
    Not evaluated
    Preview load window
    -
    Post-promotion verification
    -
    Breach rule
    -
    Recovery policy
    Revert selector + verify with live requests
    Rollback duration
    Not triggered
    Model package and image provenance