# kubectl scale deployment model-server-canary -n ai-serving –replicas=0
deployment.apps/model-server-canary scaled

# 更新主部署版本
# kubectl set image deployment/model-server model-server=fgedu/model-server:v1.3.0 -n ai-serving
deployment.apps/model-server image updated

# 验证部署状态
# kubectl rollout status deployment/model-server -n ai-serving
Waiting for deployment “model-server” rollout to finish: 2 of 3 updated replicas are available…
deployment “model-server” successfully rolled out.

# 回滚部署（如需要）
# kubectl rollout undo deployment/model-server -n ai-serving
deployment.apps/model-server rolled back

# 查看部署历史
# kubectl rollout history deployment/model-server -n ai-serving
deployment.apps/model-server
REVISION CHANGE-CAUSE
1
2
3
4 Update to v1.3.0

Model Changes:
————–
Date Action User Details
2026-04-03 Deploy v5 fengge Performance improvement
2026-04-01 Train v5 ai-pipeline Auto-triggered
2026-03-28 Alert system Data drift detected
2026-03-15 Deploy v3 fengge Scheduled release

Access Log:
———–
Date User Action IP Address
2026-04-03 fengge Predict 192.168.1.100
2026-04-03 api-service Predict 10.0.0.50
2026-04-03 batch-job Predict 10.0.0.51

Compliance Checks:
——————
Date Check Status Details
2026-04-03 Bias Detection PASS No significant bias found
2026-04-03 Privacy Check PASS No PII in predictions
2026-04-03 Explainability PASS SHAP values available
2026-04-03 Model Card PASS Documentation complete

Demographic Parity:
——————-
Group Positive Rate Disparity
Male 45.2% Baseline
Female 44.8% -0.4% (Acceptable)
Age 18-25 46.1% +0.9% (Acceptable)
Age 26-40 45.0% -0.2% (Acceptable)
Age 41-60 44.5% -0.7% (Acceptable)
Age 60+ 43.8% -1.4% (Acceptable)

Equalized Odds:
—————
Group TPR FPR Disparity
Male 92.3% 5.2% Baseline
Female 91.8% 5.5% Acceptable

Individual Fairness:
——————–
Similar instances prediction consistency: 98.5%
Fairness threshold: 95%
Status: PASS

Recommendations:
1. Continue monitoring demographic parity
2. Add fairness constraints to training
3. Document bias mitigation measures

Model Card Content:
——————-
# Model Card: fgedu_classifier v5

## Model Details
– Model Type: Classification
– Framework: TensorFlow 2.12
– Version: 5
– Created: 2026-04-03
– Author: fengge

## Intended Use
– Primary Use: Customer behavior prediction
– Users: Marketing team, Product team
– Out-of-scope: Individual decision making

## Training Data
– Source: fgedu_customer_data_v3
– Size: 1,234,567 samples
– Time Period: 2025-01 to 2026-03
– Features: 32

## Evaluation Data
– Size: 154,321 samples
– Time Period: 2026-03

## Metrics
– Accuracy: 96.78%
– Precision: 96.45%
– Recall: 96.89%
– F1-Score: 96.67%

## Ethical Considerations
– Bias Check: Passed
– Privacy: No PII used
– Explainability: SHAP available

## Limitations
– Not suitable for real-time decisions
– Requires periodic retraining
– Performance may degrade with data drift

– alert: DataDriftDetected
expr: model_psi > 0.25
for: 1m
labels:
severity: critical
annotations:
summary: “Data drift detected”
description: “Model {{ $labels.model }} PSI is {{ $value }}”

– alert: ModelErrorRateHigh
expr: rate(model_errors_total[5m]) > 0.01
for: 5m
labels:
severity: critical
annotations:
summary: “Model error rate high”
description: “Model {{ $labels.model }} error rate is {{ $value }}”

MODEL_NAME=”fgedu_classifier”
THRESHOLD_ACCURACY=0.95
THRESHOLD_LATENCY=50
THRESHOLD_ERROR_RATE=0.01

check_model_health() {
echo “Checking model health: $MODEL_NAME”
echo “================================”

ACCURACY=$(curl -s http://prometheus:9090/api/v1/query?query=model_accuracy{model=\”$MODEL_NAME\”} | jq -r ‘.data.result[0].value[1]’)
LATENCY=$(curl -s http://prometheus:9090/api/v1/query?query=model_latency_p99{model=\”$MODEL_NAME\”} | jq -r ‘.data.result[0].value[1]’)
ERROR_RATE=$(curl -s http://prometheus:9090/api/v1/query?query=rate(model_errors_total{model=\”$MODEL_NAME\”}[5m]) | jq -r ‘.data.result[0].value[1]’)

echo “Accuracy: $ACCURACY (Threshold: $THRESHOLD_ACCURACY)”
echo “Latency: ${LATENCY}s (Threshold: ${THRESHOLD_LATENCY}ms)”
echo “Error Rate: $ERROR_RATE (Threshold: $THRESHOLD_ERROR_RATE)”

if (( $(echo “$ACCURACY < $THRESHOLD_ACCURACY" | bc -l) )); then echo "WARNING: Accuracy below threshold!" send_alert "Model accuracy low: $ACCURACY" fi if (( $(echo "$LATENCY > $THRESHOLD_LATENCY” | bc -l) )); then
echo “WARNING: Latency above threshold!”
send_alert “Model latency high: ${LATENCY}s”
fi

if (( $(echo “$ERROR_RATE > $THRESHOLD_ERROR_RATE” | bc -l) )); then
echo “CRITICAL: Error rate above threshold!”
send_alert “Model error rate high: $ERROR_RATE”
fi
}

send_alert() {
curl -X POST -H ‘Content-type: application/json’ \
–data “{\”text\”:\”AI Model Alert: $1\”}” \
https://hooks.slack.com/services/xxx/yyy/zzz
}

check_model_health