Pipelines & monitoring
Pipelines execute step-by-step server-side; monitoring evaluates your alert thresholds hourly.
Pipeline stepslink
- check_circlebq_query — runs the step’s SQL against BigQuery (scan-capped) and records the row count.
- check_circletraining — queues a real Fine-tune job when the step config is a gs:// JSONL dataset.
- check_circlealert — writes a real alert event to your Studio alert feed.
- check_circlefeature_sync — registers or refreshes a feature-group entry in the Feature Store.
- check_circleevaluation / deploy — not yet backed by an execution engine; runs record these as not-executed with the reason (never a fake success).
Run limitslink
Pipeline runs are capped at 25 dispatched runs per account per rolling 24 hours (they execute real BigQuery and training work). A run over the cap fails immediately with a clear limit message and dispatches nothing.
Monitor alertslink
Enable an alert config with drift/accuracy thresholds on a monitored model; an hourly evaluator writes an alert event whenever a threshold is breached (at most once per hour per condition).