Add teardown script for POC cost-incurring resources

[shlbatra]

Summary

Adds scripts/destroy_resources.sh to tear down the three always-on resources that drive the ongoing cost of these POC projects:

• Cloud Composer 2 environment ml-pipelines-composer (GKE + Airflow, 24/7 — the biggest cost) plus its leftover DAGs bucket
• Vertex AI Feature Store — feature view iris_features then online store ml_online_store (Bigtable node billed per node-hour)
• Artifact Registry repo sahil-experiment-docker-images and all images

It mirrors the existing setup_* scripts so everything can be recreated later. BigQuery, GCS (sb-vertex), Pub/Sub and the kfp-mlops@ service account are intentionally left untouched, since the setup scripts depend on them.

Behavior / safety

• Deletes the feature view before the online store (dependency order)
• Captures the Composer DAGs bucket before deleting the env, then removes it
• Idempotent — skips resources that are already gone
• Flags: --dry-run, --yes, --skip-composer, --skip-feature-store, --skip-artifact-registry, --keep-bucket
• Confirmation prompt (type destroy) unless --yes

Recreate path

1. ./scripts/setup_artifact_registry.sh → push to main / re-run CI to rebuild images
2. ./scripts/setup_feature_store.sh
3. ./scripts/setup_composer.sh → ./scripts/sync_dags.sh

Note: setup_composer.sh pins composer-2.17.3-airflow-2.10.5; that image version may need bumping if recreated much later.

Testing

• bash -n syntax check passes
• --dry-run verified end-to-end (prints delete commands, deletes nothing)

🤖 Generated with Claude Code