المراقبة بعد النشر

import json
import logging
import time

pred_log = logging.getLogger("predictions")
pred_log.addHandler(logging.FileHandler("predictions.jsonl"))
pred_log.setLevel(logging.INFO)

def log_prediction(features: dict, price_k: float, started: float, version: str):
    pred_log.info(json.dumps({
        "time": time.time(),
        "latency_ms": round((time.time() - started) * 1000, 1),
        "model_version": version,
        "input": features,
        "output": price_k,
    }))

# inside your /predict endpoint:
#   started = time.time()
#   ... predict ...
#   log_prediction(apt.model_dump(), price, started, MODEL_VERSION)

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from scipy.stats import ks_2samp

train_area = pd.read_csv("apartments.csv")["area"]

rng = np.random.default_rng(3)
# Simulated production requests: a new compound with bigger flats opened
live_area = np.concatenate([rng.integers(60, 221, 300), rng.integers(200, 320, 150)])

stat, p_value = ks_2samp(train_area, live_area)
print(f"train mean area: {train_area.mean():.0f} m2")
print(f"live mean area:  {live_area.mean():.0f} m2")
print(f"KS p-value: {p_value:.2e}")
print("DRIFT DETECTED" if p_value < 0.01 else "no significant drift")
print(f"requests outside training range: {(live_area > train_area.max()).mean():.0%}")

plt.figure(figsize=(7, 3.4))
plt.hist(train_area, bins=30, alpha=0.6, density=True, label="training data", color="#3776AB")
plt.hist(live_area, bins=30, alpha=0.6, density=True, label="this month", color="#E0A800")
plt.xlabel("area (m²)"); plt.title("Input drift: area")
plt.legend()
plt.show()