GCP order-9 search cell - 2026-09-14 (mldp6)

First live Magento search proof on MageLift infrastructure: reindex exit 0, storefront query hits, pod-delete recycle with reconnect and no manual reindex, on a one-replica OpenSearch StatefulSet (opensearchproject/opensearch:3, 10 GB disk) behind a preview Autopilot stack. The first reindex exposed a real config defect (#env() literals under env.php system/ never resolve); the fix shipped as CONFIG__DEFAULT__ bindings, deployed via ephemeral provider v0.0.0-dialproof.13 (consumed and deleted), and the cell passed against the fixed stack. Stack retained under KEEP for the manual proof, then destroyed via RESUME teardown. Do not reuse mldp6 for profile preview (WIF pool tombstoned 30d).

Scope

Field	Value
Project	GCP acceptance project (accountRef redacted in sealed JSONL)
Region	europe-west1
Name	mldp6 (pools ml-mldp6-preview tombstoned)
Profile	preview, KEEP then RESUME teardown (pinned runId across both)
Runtime	gke-autopilot
Provider	subprocess magelift-provider-gcp v0.0.0-dialproof.13 for the fix redeploy (.12 for the base stack)
Catalog	scripts/acceptance/cells-gcp-preview.txt (13 cells) plus the manual search proof below
Artifact A	…/magento-249-rc1-static-owned-dbhost-20260816@sha256:8588b13f…fdb2be4
Seed	magento-249-sanitized-definer-free.sql.gz regenerated via setup:install of Artifact A (sha256 a2cd333c…, zero DEFINERs, values not logged)
KEEP	retained for the proof with this written reason, then destroyed

Defect and fix

indexer:reindex catalogsearch_fulltext failed pinging http://#env(MAGENTO_DC_CATALOG__SEARCH__OPENSEARCH_SERVER_HOSTNAME, …)…: config:show returned the raw literal, proving #env() under env.php's system/ section never resolves (top-level deployment sections such as db/ do). Fix (commit 6c53304): emit the same resolved search values through stock Magento's CONFIG__DEFAULT__ environment contract, which takes precedence over env.php system values — cloud adapters via CoreEnvBindings, localdev compose likewise. Unit plus local-gates green before the live redeploy.

Redeploy preview showed 3 updates (web plus cron Deployments); deploy --infra-only applied them (full deploy on a live stack is refused by the migrate Job re-create: Unauthorized). After the roll, config:show returned host mldp6-preview-search, engine opensearch.

Proof

Against the fixed stack, in ADR 0012 order:

1. Reindex exit 0: Catalog Search index has been rebuilt successfully in 00:00:01 (proof product ML-SEARCH-PROBE-1 created via REST; indexers already Update by Schedule).
2. Storefront query GET /catalogsearch/result/?q=Probe: HTTP 200, product name present (43,899 bytes).
3. Recycle: kubectl delete pod mldp6-preview-search-0, Ready in ~44s; re-query with no reindex: HTTP 200, product name present (43,860 bytes).

Probe admin credentials and API token were generated on the operator box, passed via files and stdin, never logged, and shredded after the proof.

Catalog

The 13 preview cells passed on the base stack before the proof (deploy:candidate Magento HTTP 200, CLI 2.4.9). Pinned run run-20260914t190000z-700001 covers both invocations: the KEEP run recorded the cells plus a SKIP cleanup (retention), the RESUME run recorded the PASS cleanup (candidates redacted to the gcap28 mapping, SKIP record dropped, then sealed): cleanup PASS, assert_clean ok, sealed bundle runs/gcp-gke-autopilot-magento-search-mldp6-20260914.sealed.jsonl.

Spend

cost:estimate is account-free (no live Catalog prices). Paid resources (Autopilot cluster, Cloud SQL, Memorystore, LB, OpenSearch disk) lived for the base run (~50 min) plus the proof window (~2h including the .13 CI wait) plus teardown; no leftovers.

Leftover

Order-10 preview loop rehearsal runs later in this packed session; the order-9 AWS provisioned cell attaches to the AWS session.

Non-claims

• Multi-node or HA search (single replica only)
• AOSS serverless (stays experimental)
• GKE Standard / EKS / OVH / Scaleway certified