Registry Governance and Lineage

Webhooks, approvals, audit and the chain back to training data.

0/1 done

Webhooks, approvals, lineage

Treat the registry as the control plane

Three governance practices that pay off immediately:

  1. Webhooks on transitions — fire a Slack/Jira/PagerDuty event whenever a model moves to Production. The act of shipping becomes auditable.
  2. Two-person rule — registry permissions split between proposer and approver. Same pattern as protected branches.
  3. Lineage — keep run_id, git_commit, and data_uri as tags on the version. Six months later you can answer: which data trained this model?

A real outage pattern

A model in Production started returning constant predictions. The registry alias pointed at version 14. Lineage tags showed the run used a data snapshot from a partition that had been dropped two days earlier. Rollback to version 12 took 30 sec.

Without lineage that diagnosis takes a day, not 30 seconds.

Analogy

The registry is your air-traffic control tower: nothing lands in production without a tower-cleared callsign. Webhooks are the flight strips moving across the desk — every transition leaves a paper trail.

Reflect

Map your current promotion process.

  • Who can move a model to Production today?
  • Where is that recorded? Slack? Nowhere?
  • If a regulator asked which data trained version 7, could you answer in under five minutes?

Reading in progress · 0 of 1 activity done