Switch batch inference to Feature Store offline store

[shlbatra]

Summary

• Read from canonical feature table (iris_features) instead of raw BQ tables (iris / iris_pubsub_data)
• Filter to source = 'batch_input' server-side in SQL — only scores unlabeled inference data
• Remove the conditional column rename hack (old lines 36-48) — canonical table has consistent names regardless of data source
• Use canonical feature column names (sepal_length_cm, etc.) matching the retrained model
• Use BigQueryClient(project=project_id) for query job permissions (same pattern as Step 6)

What was removed

The old inference component had a brittle if bq_table == "iris_pubsub_data" branch that renamed snake_case Pub/Sub columns to CamelCase. With the feature store, all data flows through ingest.py into iris_features with canonical names — no conditional renaming needed.

Prerequisites

• ingest.py must be run to populate iris_features with both training and batch_input rows
• bq_dataloader.py --generate-random N must be run first to create batch_input data
• Model must be retrained on canonical column names (Step 6 PR)

Test plan

• Run bq_dataloader.py --generate-random 20 then ingest.py to populate feature table
• Submit inference pipeline and confirm it reads only batch_input rows
• Verify predictions are written to iris_predictions table
• Confirm no column rename errors — canonical names used end-to-end

🤖 Generated with Claude Code

[shlbatra]

Successful run -

https://console.cloud.google.com/agent-platform/pipelines/locations/us-central1/runs/pipeline-iris-staging-inference-20260616115218?project=deeplearning-sahil

Data loaded (50 rows) -

/gcs/sb-vertex/staging/pipeline_root/57434141298/pipeline-iris-staging-inference-20260616115218/get-model_4592835321065897984/latest_model