Criando um pipeline de alertas de deriva com Evidently AI
Utilize a biblioteca Evidently para gerar um relatório HTML de desvio de dados, integrar a verificação a um script agendado e disparar um alerta por e-mail ao detectar um desvio.
Criando um pipeline de alertas de deriva com Evidently AI é uma aula grátis de Machine Learning Academy no CoddyKit. Esta é a aula 4 de 4. Você pode ler a aula completa abaixo gratuitamente — depois pratica ao vivo no navegador com um editor de código integrado e um tutor de IA 24/7. Faz parte do caminho de aprendizado de Machine Learning Academy, e seu progresso é sincronizado entre a web e o app CoddyKit. O curso de Machine Learning Academy inclui 4 aulas no total.
Partes desta aula ainda não foram traduzidas e aparecem em inglês.
What Is Evidently AI?
Evidently AI is an open-source Python library for evaluating and monitoring machine learning models in production. It provides pre-built reports and test suites covering data drift, data quality, model performance, and prediction drift. With a few lines of code, Evidently generates interactive HTML reports or JSON summaries that can be embedded in monitoring dashboards or sent as email attachments, making it the fastest way to add comprehensive monitoring to an existing ML pipeline.
# Install Evidently
# pip install evidently
import evidently
print('Evidently version:', evidently.__version__)
# Evidently organises monitoring into:
# 1. Reports -- exploratory HTML dashboards
# 2. Test Suites -- pass/fail checks for CI/CD integration
# 3. Metrics -- individual measurements (DataDriftPreset, etc.)
from evidently.report import Report
from evidently.test_suite import TestSuite
from evidently.metric_preset import DataDriftPreset, DataQualityPreset
print('Imports successful.')Preparing Reference and Current Datasets
Evidently compares a reference dataset (typically the training data or a recent stable production period) against a current dataset (the most recent production batch). Both must be pandas DataFrames with the same column schema. The reference establishes the baseline; the current is checked for drift relative to that baseline. Include the target column if ground-truth labels are available; Evidently will use them for performance metrics.
import pandas as pd
import numpy as np
from sklearn.datasets import make_classification
np.random.seed(42)
# Simulate training data (reference)
X_ref, y_ref = make_classification(n_samples=5000, n_features=8, random_state=42)
reference = pd.DataFrame(X_ref, columns=[f'feature_{i}' for i in range(8)])
reference['target'] = y_ref
# Simulate production batch with some drift
X_cur = X_ref[:1000].copy()
X_cur[:, 2] += 2.0 # shift feature_2
X_cur[:, 5] *= 1.8 # scale feature_5
current = pd.DataFrame(X_cur, columns=[f'feature_{i}' for i in range(8)])
current['target'] = y_ref[:1000] # labels may or may not be available
print('Reference shape:', reference.shape)
print('Current shape:', current.shape)Generating a Data Drift Report
The DataDriftPreset generates a comprehensive drift report covering all features. It automatically selects the appropriate statistical test for each column (KS test for continuous features, chi-squared for categorical) and reports the drift score, p-value, and a yes/no drift decision for each. The HTML output includes interactive visualisations showing the reference vs current distribution for each drifted feature.
from evidently.report import Report
from evidently.metric_preset import DataDriftPreset
# Build and run the report
data_drift_report = Report(metrics=[DataDriftPreset()])
data_drift_report.run(
reference_data=reference,
current_data=current
)
# Save to HTML file
data_drift_report.save_html('/tmp/data_drift_report.html')
print('Report saved to /tmp/data_drift_report.html')
# Extract as dictionary for programmatic use
result_dict = data_drift_report.as_dict()
drift_summary = result_dict['metrics'][0]['result']
print('Dataset drift detected:', drift_summary['dataset_drift'])
print('Features drifted:', drift_summary['number_of_drifted_columns'])Column Drift Details
Access per-column drift results programmatically to build automated alerts. Evidently returns a dictionary with each column's drift score, statistical test used, p-value, and a boolean drift flag. Filter for drifted columns and include them in your alert message to help the team immediately understand which features have changed and by how much.
from evidently.report import Report
from evidently.metric_preset import DataDriftPreset
report = Report(metrics=[DataDriftPreset()])
report.run(reference_data=reference, current_data=current)
result = report.as_dict()['metrics'][0]['result']
print('Per-column drift summary:')
for col, info in result['drift_by_columns'].items():
drift_flag = info.get('drift_detected', False)
stat_name = info.get('stattest_name', 'N/A')
p_val = info.get('p_value', 'N/A')
status = '*** DRIFTED ***' if drift_flag else 'stable'
print(f' {col:12s}: {stat_name} p={p_val:.4f} [{status}]' if isinstance(p_val, float)
else f' {col}: {status}')Data Quality Report
Beyond drift, Evidently's DataQualityPreset checks for data quality issues in the current batch: missing values, constant columns, duplicates, and out-of-range values. These issues are often the first sign of an upstream data pipeline failure rather than true drift. Running data quality checks before drift checks prevents false drift alarms caused by missing imputation or schema changes in the ETL pipeline.
from evidently.report import Report
from evidently.metric_preset import DataQualityPreset
import pandas as pd
import numpy as np
# Introduce some quality issues
current_with_issues = current.copy()
current_with_issues.loc[50:80, 'feature_0'] = np.nan # missing values
current_with_issues.loc[100:110, 'feature_3'] = 99999 # outliers
quality_report = Report(metrics=[DataQualityPreset()])
quality_report.run(
reference_data=reference,
current_data=current_with_issues
)
quality_report.save_html('/tmp/data_quality_report.html')
quality_dict = quality_report.as_dict()
print('Quality report generated with', len(quality_dict['metrics']), 'metrics.')Test Suites for CI/CD: Pass or Fail
Reports are for humans; test suites are for machines. A test suite runs a battery of configurable checks and returns a pass/fail result that can gate pipeline steps. If any test fails, the pipeline raises an exception, the CI/CD step fails, and the alert fires. Common tests: number of drifted columns below threshold, missing value rate below threshold, all columns within expected value ranges.
from evidently.test_suite import TestSuite
from evidently.tests import (
TestNumberOfDriftedColumns,
TestShareOfMissingValues,
TestNumberOfConstantColumns
)
test_suite = TestSuite(tests=[
TestNumberOfDriftedColumns(lt=2), # fail if > 1 column drifts
TestShareOfMissingValues(lt=0.01), # fail if >1% missing
TestNumberOfConstantColumns(lt=1) # fail if any constant column
])
test_suite.run(reference_data=reference, current_data=current)
test_suite.save_html('/tmp/test_suite_report.html')
result = test_suite.as_dict()
summary = result['summary']
print(f'Tests passed: {summary["passed_tests"]}/{summary["total_tests"]}')
print('Overall status:', 'PASS' if summary['all_passed'] else 'FAIL')Integrating Evidently into a Scheduled Script
Automate drift monitoring by running Evidently on a daily cron job. The script loads the previous day's production batch, compares it to the reference dataset, runs the test suite, and sends an alert if tests fail. This pattern requires no ML expertise to maintain — the scheduled script runs silently when everything is healthy and raises an alarm only when action is needed.
import pandas as pd
from evidently.test_suite import TestSuite
from evidently.tests import TestNumberOfDriftedColumns, TestShareOfMissingValues
from datetime import date
def daily_drift_check(reference_path, production_path, alert_threshold=2):
reference = pd.read_parquet(reference_path)
current = pd.read_parquet(production_path)
suite = TestSuite(tests=[
TestNumberOfDriftedColumns(lt=alert_threshold),
TestShareOfMissingValues(lt=0.05)
])
suite.run(reference_data=reference, current_data=current)
result = suite.as_dict()
passed = result['summary']['all_passed']
report_path = f'/tmp/drift_{date.today()}.html'
suite.save_html(report_path)
return passed, report_path, result['summary']
# In cron script:
# passed, report, summary = daily_drift_check('data/reference.parquet', 'data/today.parquet')
# if not passed:
# send_email_alert(report, summary)
print('daily_drift_check function ready.')Sending Email Alerts on Drift Detection
When the test suite fails, send an email alert with the HTML report attached. Use Python's built-in smtplib and email modules to compose a message with the drift summary in the body and the full Evidently HTML report as an attachment. The recipient can open the report to see exactly which features drifted and by how much, without needing access to the production environment.
import smtplib
import os
from email.mime.multipart import MIMEMultipart
from email.mime.text import MIMEText
from email.mime.application import MIMEApplication
def send_drift_alert(recipient, report_path, summary):
smtp_host = os.getenv('SMTP_HOST', 'smtp.gmail.com')
smtp_user = os.getenv('SMTP_USER')
smtp_pass = os.getenv('SMTP_PASS')
msg = MIMEMultipart()
msg['From'] = smtp_user
msg['To'] = recipient
msg['Subject'] = '[ML ALERT] Data drift detected in production'
body = (f'Drift alert triggered.\n'
f'Tests passed: {summary["passed_tests"]}/{summary["total_tests"]}\n'
f'See attached HTML report for details.')
msg.attach(MIMEText(body, 'plain'))
with open(report_path, 'rb') as f:
attachment = MIMEApplication(f.read(), _subtype='html')
attachment.add_header('Content-Disposition', 'attachment',
filename='drift_report.html')
msg.attach(attachment)
print(f'Alert email sent to {recipient}')
print('send_drift_alert function ready.')Slack Webhook Integration
For teams that use Slack, posting a concise drift alert to a dedicated #ml-monitoring channel is more immediately visible than email. Include the number of drifted features, the most severely drifted column, and a link to the saved HTML report in cloud storage (S3 or GCS). Use Slack Block Kit for formatted messages with colour-coded status (green = healthy, red = drift detected).
import requests
import json
import os
def post_slack_alert(summary, drifted_columns, report_url):
webhook_url = os.getenv('SLACK_WEBHOOK_URL')
if not webhook_url:
print('No Slack webhook configured.')
return
all_passed = summary['all_passed']
icon = ':white_check_mark:' if all_passed else ':rotating_light:'
color = 'good' if all_passed else 'danger'
status_text = 'All checks passed' if all_passed else 'DRIFT DETECTED'
message = {
'attachments': [{
'color': color,
'title': f'{icon} ML Monitoring Alert',
'text': (f'Status: {status_text}\n'
f'Drifted features: {drifted_columns}\n'
f'Tests: {summary["passed_tests"]}/{summary["total_tests"]} passed\n'
f'<{report_url}|View full report>'),
'fallback': status_text
}]
}
response = requests.post(webhook_url, data=json.dumps(message))
print('Slack response:', response.status_code)Saving Reports to S3 for Audit Trail
Store every generated Evidently report in S3 or GCS with a date-stamped key to create a permanent audit trail. This enables retrospective analysis: when a model's performance drops in April, retrieve the March drift report to see whether drift was already visible weeks earlier. Configure an S3 lifecycle policy to retain reports for 2 years and automatically archive older ones to Glacier storage for cost efficiency.
import boto3
import os
from datetime import date
def upload_report_to_s3(local_path, bucket, model_name):
s3 = boto3.client(
's3',
aws_access_key_id=os.getenv('AWS_ACCESS_KEY_ID'),
aws_secret_access_key=os.getenv('AWS_SECRET_ACCESS_KEY')
)
today = date.today().isoformat()
s3_key = f'ml-monitoring/{model_name}/{today}/drift_report.html'
s3.upload_file(
local_path,
bucket,
s3_key,
ExtraArgs={'ContentType': 'text/html'}
)
report_url = f'https://{bucket}.s3.amazonaws.com/{s3_key}'
print(f'Report uploaded: {report_url}')
return report_url
print('upload_report_to_s3 function ready.')Dashboard Integration with Grafana
Evidently produces HTML reports suitable for human review, but for operational monitoring you need a real-time dashboard that updates automatically. Extract PSI and KS statistics as numeric values from Evidently test results, push them to Prometheus (via the prometheus_client library), and visualise in Grafana. This gives your SRE team the same observability tooling for ML models as for application services, enabling drift alerts alongside latency and error-rate alerts in one place.
# Export Evidently metrics to Prometheus
from prometheus_client import Gauge, start_http_server
import time
# Define Prometheus gauges for drift metrics
psi_gauge = Gauge('ml_feature_psi', 'Population Stability Index', ['feature_name'])
ks_gauge = Gauge('ml_feature_ks_statistic', 'KS test statistic', ['feature_name'])
# Push computed drift values (call after running Evidently tests)
def push_drift_metrics(drift_results: dict):
for feature, metrics in drift_results.items():
psi_gauge.labels(feature_name=feature).set(metrics.get('psi', 0))
ks_gauge.labels(feature_name=feature).set(metrics.get('ks_stat', 0))
# start_http_server(8000) # Prometheus scrapes this port
print('Expose metrics at :8000/metrics -> scrape in Prometheus -> display in Grafana')Quick Check
Test your understanding of Machine Learning with Python concepts from this lesson.
Lesson Recap
In this lesson you learned: Evidently AI generates comprehensive drift and data quality reports with a few lines of Python, Test Suites provide pass/fail results that integrate directly into CI/CD pipelines and scheduled cron jobs, and combining S3 storage, email, and Slack notifications creates a complete automated alert pipeline that catches drift before it visibly impacts users. Next up we enter the Responsible AI section, starting with SHAP values for interpreting global and local feature importance in complex models.
Perguntas Frequentes
A aula “Criando um pipeline de alertas de deriva com Evidently AI” é grátis?
Sim — o texto completo de “Criando um pipeline de alertas de deriva com Evidently AI” é grátis para ler aqui na web. Para praticá-la interativamente (um editor de código integrado e um tutor de IA 24/7) e desbloquear o restante do curso de Machine Learning Academy, atualize para CoddyKit PRO. O curso de Machine Learning Academy inclui 4 aulas no total.
O que vou aprender em “Criando um pipeline de alertas de deriva com Evidently AI”?
Utilize a biblioteca Evidently para gerar um relatório HTML de desvio de dados, integrar a verificação a um script agendado e disparar um alerta por e-mail ao detectar um desvio. Você pratica Machine Learning Academy com código prático que executa diretamente no navegador, e um tutor de IA 24/7 responde suas dúvidas enquanto trabalha na aula.
Preciso ter experiência prévia para começar Machine Learning Academy?
Nenhuma experiência prévia é necessária. Machine Learning Academy no CoddyKit é estruturado para alunos iniciantes até avançados, então você pode começar aqui ou desde o início e aprender no seu ritmo. Esta é a aula 4 de 4.
Quanto tempo leva a aula “Criando um pipeline de alertas de deriva com Evidently AI”?
A maioria das aulas CoddyKit leva cerca de 5–10 minutos. Cada uma é compacta e interativa, então você faz progresso constante e retoma exatamente de onde parou entre web e app.
Posso escrever e executar código nesta aula de Machine Learning Academy?
Sim. Cada aula de Machine Learning Academy inclui um editor de código integrado, então você escreve e executa código real direto no navegador e recebe feedback de IA instantaneamente — nenhuma configuração local necessária.
Todas as aulas deste curso
- Deriva dos dados: mudanças na distribuição das características ao longo do tempo
- Deriva de conceito: quando a relação entre X e Y muda
- Monitoramento das distribuições de previsões e das pontuações de confiança
- Criando um pipeline de alertas de deriva com Evidently AI