监控预测：记录输入与输出
您将为 API 添加预测记录，讨论数据漂移和模型过时问题，并构思重新训练触发器和部署管道。
监控预测：记录输入与输出 是 CoddyKit 上的免费 Machine Learning Academy 课时。 这是第 4 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 Machine Learning Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Machine Learning Academy 课程共包含 4 节课。
本课时的部分内容尚未翻译，以英文显示。
Why Monitor a Deployed Model?
A model that performed excellently at deployment can silently degrade as the real world changes. Customer behaviour shifts, new product categories appear, sensors drift, economic conditions change — all causing the input distribution to diverge from training data. Without monitoring, you discover model failure when users complain or business metrics drop. Prediction monitoring catches degradation early, before it impacts users.
What to Log: The Three W's
Every prediction event should log: When (timestamp), What (input features and the model's output/prediction/probability), and ideally Whether (the eventual ground truth label, once available). The input features are needed for data drift detection; the outputs enable confidence monitoring; the ground truth enables accuracy monitoring over time.
Adding Logging to the FastAPI Endpoint
Extend the FastAPI prediction route to log each request to a structured JSONL file (JSON Lines — one JSON object per line). JSONL is easy to append to, easy to parse, and works well with tools like pandas, Spark, and cloud log aggregators. Each log line is one prediction event.
import json
from datetime import datetime
from fastapi import FastAPI
from pydantic import BaseModel
import numpy as np
app = FastAPI()
LOG_FILE = '/tmp/prediction_log.jsonl'
def log_prediction(features: dict, prediction: int, confidence: float):
entry = {
'timestamp': datetime.utcnow().isoformat(),
'features': features,
'predicted_class': prediction,
'confidence': confidence
}
with open(LOG_FILE, 'a') as f:
f.write(json.dumps(entry) + '\n')
# Call inside the predict route:
# log_prediction(features.dict(), pred, confidence)
print('Logging enabled — entries written to', LOG_FILE)Simulating and Reading the Log
After accumulating predictions, read the JSONL file into a pandas DataFrame for analysis. Each row is a prediction event. You can compute statistics over any time window: rolling mean confidence, feature distribution shifts, and prediction class frequency.
import json
import pandas as pd
from datetime import datetime
import random
# Simulate log entries
LOG_FILE = '/tmp/prediction_log.jsonl'
with open(LOG_FILE, 'w') as f:
for i in range(100):
entry = {
'timestamp': datetime.utcnow().isoformat(),
'features': {'sepal_length': round(4.5 + random.gauss(1, 0.5), 2),
'sepal_width': round(3.0 + random.gauss(0, 0.3), 2),
'petal_length': round(1.0 + random.gauss(2, 1), 2),
'petal_width': round(0.2 + random.gauss(0.5, 0.2), 2)},
'predicted_class': random.choice([0, 1, 2]),
'confidence': round(random.uniform(0.6, 1.0), 4)
}
f.write(json.dumps(entry) + '\n')
# Read back as DataFrame
rows = [json.loads(line) for line in open(LOG_FILE)]
df = pd.json_normalize(rows)
print(df.shape)
print(df[['confidence', 'predicted_class']].describe())Monitoring Confidence Over Time
Plot rolling mean confidence over time. A downward trend signals that the model is becoming less certain about its predictions — a classic early warning sign of data drift. Set a threshold (e.g., rolling confidence below 0.7) and trigger an alert. This gives you a warning before accuracy actually drops.
import pandas as pd
import matplotlib.pyplot as plt
import json
df = pd.json_normalize([json.loads(l) for l in open('/tmp/prediction_log.jsonl')])
df['timestamp'] = pd.to_datetime(df['timestamp'])
df = df.sort_values('timestamp').reset_index(drop=True)
# Rolling window of 20 predictions
df['rolling_confidence'] = df['confidence'].rolling(20).mean()
plt.figure(figsize=(10, 4))
plt.plot(df.index, df['confidence'], alpha=0.3, label='Per-prediction')
plt.plot(df.index, df['rolling_confidence'], linewidth=2, label='Rolling mean (n=20)')
plt.axhline(0.7, color='red', linestyle='--', label='Alert threshold')
plt.xlabel('Prediction number')
plt.ylabel('Confidence')
plt.title('Prediction Confidence Over Time')
plt.legend()
plt.show()Detecting Feature Distribution Shifts
Compare the distribution of each input feature in recent predictions against the training distribution. The Population Stability Index (PSI) is a common metric: PSI < 0.1 means no significant shift; 0.1–0.2 is moderate drift; >0.2 is severe drift requiring retraining. Alternatively, a Kolmogorov-Smirnov test tests whether two samples come from the same distribution.
import numpy as np
from scipy import stats
import pandas as pd
import json
df = pd.json_normalize([json.loads(l) for l in open('/tmp/prediction_log.jsonl')])
# Simulated training distribution for sepal_length
train_sepal = np.random.normal(5.8, 0.8, 500) # reference
recent_sepal = df['features.sepal_length'].values
# KS test: tests if two samples come from same distribution
stat, p_value = stats.ks_2samp(train_sepal, recent_sepal)
print(f'KS statistic: {stat:.4f} p-value: {p_value:.4f}')
if p_value < 0.05:
print('WARNING: sepal_length distribution has drifted!')
else:
print('No significant drift detected in sepal_length')Monitoring Prediction Class Frequencies
Plot the rolling frequency of each predicted class. If class 0 was predicted 60% of the time during training-period evaluation but is now being predicted only 10% of the time, something in the input data has shifted. Class distribution shifts are often easier to detect than subtle feature distribution changes and can serve as a quick first alert.
import pandas as pd
import json
df = pd.json_normalize([json.loads(l) for l in open('/tmp/prediction_log.jsonl')])
# Compute rolling class frequencies in windows of 20
window = 20
rolling_freq = pd.get_dummies(df['predicted_class']).rolling(window).mean()
rolling_freq.columns = [f'class_{c}' for c in rolling_freq.columns]
print('Recent class frequencies (last 20 predictions):')
print(rolling_freq.tail(1).to_string())
print('\nTraining class frequency (expected):')
print('class_0: 0.33 class_1: 0.33 class_2: 0.33')Structured Logging for Production
In production, write logs to a structured logging system rather than a flat file. Options include: Python logging module with JSON formatter, cloud logging services (AWS CloudWatch, GCP Cloud Logging, Datadog), or a database. Structured logs are queryable, scalable, and integrate with dashboards like Grafana for real-time monitoring.
import logging
import json
from datetime import datetime
# Configure structured JSON logging
class JSONFormatter(logging.Formatter):
def format(self, record):
log_data = {
'time': datetime.utcnow().isoformat(),
'level': record.levelname,
'message': record.getMessage()
}
if hasattr(record, 'prediction_data'):
log_data.update(record.prediction_data)
return json.dumps(log_data)
logger = logging.getLogger('ml_predictor')
handler = logging.StreamHandler()
handler.setFormatter(JSONFormatter())
logger.addHandler(handler)
logger.setLevel(logging.INFO)
# Usage in prediction route:
extra = {'prediction_data': {'features': {'x': 1.5}, 'pred': 0, 'conf': 0.95}}
logger.info('prediction', extra=extra)Data Drift vs Concept Drift
Two distinct failure modes exist: Data drift — the distribution of input features changes (e.g., a sensor recalibrates, new customer demographics emerge). Concept drift — the relationship between inputs and the target changes (e.g., what constitutes 'fraud' evolves as fraudsters adapt). Data drift can sometimes be handled by retraining on fresh data; concept drift requires rethinking features or the modelling approach.
Retraining Triggers
Define clear retraining triggers before deployment. Common triggers: scheduled (retrain every 30 days regardless), performance-based (retrain if AUC on a labelled holdout drops below a threshold), or drift-based (retrain when PSI exceeds 0.2 on any key feature). Automating the trigger and the retraining pipeline prevents the model from silently degrading.
from datetime import datetime, timedelta
class RetrainingPolicy:
def __init__(self, last_trained: datetime, max_age_days=30,
min_auc=0.85, max_drift_psi=0.2):
self.last_trained = last_trained
self.max_age_days = max_age_days
self.min_auc = min_auc
self.max_drift_psi = max_drift_psi
def should_retrain(self, current_auc=None, max_psi=None):
age = (datetime.utcnow() - self.last_trained).days
if age >= self.max_age_days:
return True, f'Model is {age} days old (max {self.max_age_days})'
if current_auc and current_auc < self.min_auc:
return True, f'AUC {current_auc:.3f} below threshold {self.min_auc}'
if max_psi and max_psi > self.max_drift_psi:
return True, f'PSI {max_psi:.3f} above threshold {self.max_drift_psi}'
return False, 'No trigger met — model healthy'
policy = RetrainingPolicy(last_trained=datetime.utcnow() - timedelta(days=35))
print(policy.should_retrain(current_auc=0.90))The Monitoring Dashboard Sketch
A production ML monitoring dashboard typically shows: rolling prediction confidence over time, class distribution over time, feature distribution histograms comparing training vs recent predictions, and (when labels are available) rolling accuracy or AUC. Tools like Grafana + Prometheus, Evidently AI, or WhyLogs provide pre-built components for ML monitoring without building from scratch.
# Evidently AI example (pip install evidently)
from evidently.report import Report
from evidently.metric_preset import DataDriftPreset
import pandas as pd
import numpy as np
# Reference: training feature distributions
reference = pd.DataFrame({
'sepal_length': np.random.normal(5.8, 0.8, 300),
'petal_length': np.random.normal(3.7, 1.7, 300)
})
# Current: recent predictions features
current = pd.DataFrame({
'sepal_length': np.random.normal(6.5, 1.2, 100), # drifted!
'petal_length': np.random.normal(3.7, 1.7, 100)
})
report = Report(metrics=[DataDriftPreset()])
report.run(reference_data=reference, current_data=current)
report.save_html('/tmp/drift_report.html')
print('Drift report saved to /tmp/drift_report.html')Quick Check
Test your understanding of prediction monitoring and logging from this lesson.
Lesson Recap
In this lesson you learned: log every prediction event with timestamp, input features, predicted class, and confidence to enable drift detection, monitor rolling confidence and feature distributions to catch degradation before it affects users, and define automated retraining triggers (schedule, performance threshold, or drift threshold) to keep the model healthy in production. This completes the Model Persistence and Deployment Basics course.
常见问题解答
「监控预测：记录输入与输出」课时是免费的吗？
是的 — 「监控预测：记录输入与输出」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Machine Learning Academy 课程的其余内容，请升级到 CoddyKit PRO。 Machine Learning Academy 课程共包含 4 节课。
「监控预测：记录输入与输出」这节课中我会学到什么？
您将为 API 添加预测记录，讨论数据漂移和模型过时问题，并构思重新训练触发器和部署管道。 你通过在浏览器中直接运行的动手代码来练习 Machine Learning Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。
学习 Machine Learning Academy 需要有经验吗？
无需任何先前经验。CoddyKit 上的 Machine Learning Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 4 节课，共 4 节。
「监控预测：记录输入与输出」课时需要多长时间？
大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。
我能在这节 Machine Learning Academy 课中编写并运行代码吗？
能。每节 Machine Learning Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。