1. Evaluate as usual
Your existing Airflow, CI, or other orchestrator runs training and model evaluation, then asks DataXPipe to execute the SQL check.
DataXPipe for ML teams
Keep training, evaluation, and orchestration in the tools you already use. DataXPipe checks that evaluation data is fresh and that a metric such as ROC AUC meets your release bound before downstream batch scoring proceeds.
Your existing Airflow, CI, or other orchestrator runs training and model evaluation, then asks DataXPipe to execute the SQL check.
Use a SQL query that returns one scalar as metric and set a minimum, maximum, or both. Invalid or missing metrics fail closed.
A failed check can revoke the target Passport. Configure lineage and a circuit breaker for the downstream batch scoring workflow to pause it.
pipeline:
name: daily_risk_model_release
description: "Validate evaluation metrics before batch scoring"
owner: ml-platform
environment: production
schedule: "0 6 * * *"
tags: [ml, batch]
sources:
- id: model_eval_metrics
type: warehouse
connection_ref: analytics
object: ML.MODEL_EVAL_METRICS
load_mode: full
targets:
- id: batch_scores
type: warehouse
connection_ref: analytics
object: ML.BATCH_SCORES
write_mode: append
passport:
enabled: true
sla_minutes: 1440
checks:
- id: evaluation_freshness
type: freshness
target: model_eval_metrics
watermark_column: evaluated_at
max_delay_minutes: 1440
- id: roc_auc_release_gate
type: ml_metric_threshold
target: model_eval_metrics
sql: "SELECT ROC_AUC AS metric FROM ML.MODEL_EVAL_METRICS ORDER BY EVALUATED_AT DESC LIMIT 1"
min_value: 0.82
lineage:
edges:
- from: model_eval_metrics
to: batch_scores
via: daily_risk_model_releaseDataXPipe validates and gates workflows; it does not train models, select acceptance thresholds, or replace your model registry. Choose thresholds from your product risk and evaluation practice.