Fix silent no-op when destroying the Vertex AI Feature Store

[shlbatra]

Problem

./scripts/destroy_resources.sh reported success but left ml_online_store alive and billing a Bigtable node.

The feature-store deletes were guarded by:

if gcloud ai feature-online-stores describe "$ONLINE_STORE_ID" ... >/dev/null 2>&1; then

gcloud ai feature-online-stores is a beta-only command group and isn't present in a stock gcloud install. gcloud exits 2 with ERROR: (gcloud.ai) Invalid choice: 'feature-online-stores', but stderr goes to /dev/null — so the guard reads "command failed" as "resource doesn't exist" and prints:

Online store 'ml_online_store' not found — skipping.
Feature Store torn down.

No delete was ever attempted. The asymmetry is that creation goes through the Python SDK (setup_feature_store.sh → python -m feature_store.setup → aiplatform_v1), which hits the REST API directly and works fine.

Fix

• Talk to the Vertex AI REST API directly (fs_get / fs_delete helpers), matching how the store is created. No dependency on the beta gcloud component.
• fs_get succeeds only on HTTP 200, so a broken call can no longer masquerade as "not found".
• fs_delete polls the returned long-running operation to completion instead of returning before the resource is actually gone.
• Delete every feature view in the store rather than only the hardcoded $FEATURE_VIEW_ID — a leftover view blocks the store delete with FAILED_PRECONDITION.

Verification

bash -n passes. Dry run against the live project now discovers both resources it previously missed:

--- Feature Store ---
Deleting feature view 'iris_features'...
  [dry-run] fs_delete .../featureOnlineStores/ml_online_store/featureViews/iris_features
Deleting online store 'ml_online_store' (stops Bigtable node billing)...
  [dry-run] fs_delete .../featureOnlineStores/ml_online_store

Not yet run for real — ml_online_store is still up.

🤖 Generated with Claude Code