Model handover checklist (DS → DE)
☐ Model saved in an agreed format and location
☐ Input feature names, types and order documented
☐ Preprocessing included in the model or documented exactly
☐ Output meaning (probability? class id?) documented
☐ Expected input ranges for validation
☐ Training data version recorded
☐ Owner and retraining schedule agreed

import tensorflow as tf

model = tf.keras.models.load_model("churn_model.keras")
model.summary()
print("Input shape:", model.input_shape)
print("Output shape:", model.output_shape)
# Confirm these match the feature table your pipeline produces

import sqlite3
import pandas as pd
import tensorflow as tf
from datetime import date

FEATURES = ["tenure_months", "monthly_spend", "support_tickets"]
model = tf.keras.models.load_model("churn_model.keras")

with sqlite3.connect("warehouse.db") as con:
    df = pd.read_sql("SELECT customer_id, " + ", ".join(FEATURES) + " FROM customer_features", con)
    assert df[FEATURES].notna().all().all(), "nulls in features"
    df["churn_score"] = model.predict(df[FEATURES].astype("float32").values,
                                      batch_size=1024, verbose=0).ravel()
    df["scored_on"] = date.today().isoformat()
    df["model_version"] = "churn_v3"
    df[["customer_id", "churn_score", "scored_on", "model_version"]] \
        .to_sql("churn_scores", con, if_exists="append", index=False)

# Example cron entry: run daily at 02:30
30 2 * * * /opt/jobs/venv/bin/python /opt/jobs/score_churn.py >> /var/log/score_churn.log 2>&1

# In Airflow, the same script becomes a PythonOperator or a
# containerised task, with retries and alerts on failure.

import pandas as pd

train = pd.read_parquet("train_features.parquet")
today = pd.read_parquet("features_today.parquet")

for col in ["tenure_months", "monthly_spend", "support_tickets"]:
    t_mean, n_mean = train[col].mean(), today[col].mean()
    t_std = train[col].std()
    shift = abs(n_mean - t_mean) / t_std if t_std else 0
    flag = "DRIFT" if shift > 0.5 else "ok"
    print(f"{col:16} train={t_mean:8.2f} today={n_mean:8.2f} shift={shift:.2f} {flag}")
# threshold 0.5 std is an example; agree thresholds with the ML team

TFX pipeline components and the DE angle
ExampleGen       → ingest data into the pipeline (DE owns sources)
StatisticsGen    → compute dataset statistics
SchemaGen        → infer an expected schema
ExampleValidator → flag anomalies against the schema
Transform        → shared feature preprocessing
Trainer / Evaluator → ML team's area
Pusher           → deploy a validated model
You do not need to run TFX to use these ideas in your own pipelines.

Batch Inference Job Spec
Model: ______  Version: ______  Location: ______
Input table: ______  Columns (order): ______
Validation: nulls ☐ ranges ☐ schema ☐ volume ☐
Schedule: ______  Max runtime: ______
Output table: ______ (append, with run date + model version)
On failure: retry ×2, then alert ______

Feature Contract
Feature           Type     Unit      Range        Source
tenure_months     int      months    0-240        crm.customers
monthly_spend     float    INR       0-500000     billing.invoices
support_tickets   int      count     0-100        support.tickets
Change policy: notify ML team 2 weeks before any change

Model data monitoring dashboard
- Rows scored per day
- Null rate per feature
- Mean and std per feature vs training
- Share of each category for categorical features
- Score distribution over time
- Job runtime and failures