Blue/green rollout

Blue/green rollout is the safest way to ship terminology changes when search quality matters.

The idea is simple:

blue  = current active snapshot
green = candidate snapshot

The green snapshot is validated before it becomes the active runtime version for a binding.

Why blue/green matters for terminology

A small alias change can improve recall or add noise. For example, expanding pg globally may help infrastructure queries but harm document-search queries where pg means page.

Blue/green rollout gives teams a way to validate before switching traffic.

Blue snapshot

The currently active runtime version pinned to a production binding.

ActiveKnown-good

Green snapshot

The candidate version produced after review, import, enrichment, or apply planning.

CandidateValidation

Alias swap

A controlled activation step that moves a binding or downstream alias from blue to green.

ActivationRollback path

Rollback

If search quality regresses, repin the binding or alias to the previous snapshot.

Safe recovery

Rollout checklist

1. Publish a candidate snapshot.
2. Run apply planning and evidence checks.
3. Run smoke searches for representative queries.
4. Compare canonical query output and matched aliases.
5. Validate enrichment job status and target field health.
6. Activate the candidate snapshot for the binding.
7. Monitor search quality and operator feedback.
8. Roll back if noise increases.

Example runtime state

{
  "binding_id": 42,
  "blue_snapshot": "infra_incidents@2026-05-29T18-00Z",
  "green_snapshot": "infra_incidents@2026-05-30T21-00Z",
  "active_snapshot": "infra_incidents@2026-05-29T18-00Z",
  "candidate_status": "validated"
}

After activation:

{
  "binding_id": 42,
  "active_snapshot": "infra_incidents@2026-05-30T21-00Z",
  "previous_snapshot": "infra_incidents@2026-05-29T18-00Z"
}

What to measure

A rollout should not be judged only by whether the job succeeded. It should be judged by whether search behavior improved.

Track at least:

• matched aliases per query;
• canonical query output;
• zero-result rate;
• noisy term expansions;
• evidence coverage;
• rollback frequency;
• representative query set results.

Practical default

For the beta stage, keep the rollout rule conservative:

Do not auto-activate high-impact terminology changes without a human review and a successful smoke check.