618 大促，我按 Prophet 预测扩了 12 个 Pod，结果被打穿了

训练数据时间范围: 2026-03-20 ~ 2026-06-17（90 天）

覆盖的促销日: ['劳动节📌']
缺失的促销日: ['618大促❌', '双十二❌', '元旦❌', '情人节❌', '三八节❌']

# 翻车时的模型参数
model = Prophet(
    changepoint_prior_scale=0.05,   # ← 默认值，适合平稳序列
)

from prophet import Prophet
from sklearn.metrics import mean_absolute_percentage_error

# 方案 1：训练数据从 90 天扩展到 365 天
qps_df = fetch_metric('qps', days=365)

# 方案 2：调大 changepoint_prior_scale
model = Prophet(
    changepoint_prior_scale=0.15,       # 0.05 → 0.15，趋势更灵活
    yearly_seasonality=True,            # 启用年度季节性
    holidays=promotions,                # 显式加入大促日
)

# 方案 3：增加回测验证
defbacktest_model(df, train_days=60, test_days=7):
"""回测验证：用历史数据检验预测精度"""
    model.fit(df.iloc[:len(df)-test_days])
    forecast = model.predict(...)
    mape = mean_absolute_percentage_error(test['y'], pred['yhat'])
assert mape < 0.15, f"回测精度不足: {mape:.2%}"

📊 修复前后对比
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
         指标        修复前      修复后
    ───────────────────────────────────
   训练数据范围       90 天      365 天
   回测 MAPE         18.5%      9.2%
   促销活动实测误差   83%        3.8% ✅
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

"""
从 Prometheus 提取时序数据，用于 Prophet 预测
"""

import requests
import pandas as pd
from datetime import datetime, timedelta

PROMETHEUS_URL = "http://localhost:9090"

deffetch_metric(metric: str, days: int = 365) -> pd.DataFrame:
"""获取指标历史数据（默认 365 天）"""
    end = datetime.now()
    start = end - timedelta(days=days)

    resp = requests.get(
f"{PROMETHEUS_URL}/api/v1/query_range",
        params={
"query": metric,
"start": start.timestamp(),
"end": end.timestamp(),
"step": "3600s"# 1 小时粒度
        }
    )

    values = resp.json()["data"]["result"][0]["values"]
    df = pd.DataFrame(values, columns=["timestamp", "value"])
    df["ds"] = pd.to_datetime(df["timestamp"], unit="s")
    df["y"] = df["value"].astype(float)

return df[["ds", "y"]]

# 提取订单服务 QPS
qps_df = fetch_metric(
'sum(rate(http_requests_total{job="order-service"}[5m]))'
)
print(f"时间范围: {qps_df['ds'].min()} ~ {qps_df['ds'].max()}")
print(f"数据行数: {len(qps_df)}")

"""
Prophet 容量预测：业务日历特征工程
"""

from prophet import Prophet

# 📅 业务日历 — 这是最重要的特征！
promotions = pd.DataFrame({
"holiday": "promotion",
"ds": pd.to_datetime([
"2025-06-18",              # 去年 618（训练数据）
"2025-11-11",              # 去年双11
"2026-06-18", "2026-06-19", # 今年 618
"2026-11-11", "2026-11-12", # 今年双11
    ]),
"lower_window": -1,   # 大促前一天就有影响
"upper_window": 1,    # 大促后一天也不放松
})

# 构建模型
model = Prophet(
    holidays=promotions,
    yearly_seasonality=True,        # 年度季节性
    weekly_seasonality=True,        # 周内季节性
    daily_seasonality=True,         # 日内季节性
    changepoint_prior_scale=0.15,   # 趋势灵活性（大促场景调大）
)

# 训练
model.fit(qps_df)

# 预测未来 7 天（168 小时）
future = model.make_future_dataframe(periods=168, freq="h")
forecast = model.predict(future)

# 输出预测结果
result = forecast[["ds", "yhat", "yhat_lower", "yhat_upper"]].tail(24)
print(result.to_string(index=False))

                 ds       yhat  yhat_lower  yhat_upper
2026-07-01 00:00:00   1250.32      980.15     1520.48
2026-07-01 08:00:00   3500.21     3100.45     3900.00
2026-07-01 12:00:00   5200.87     4700.32     5701.42
2026-07-01 20:00:00   4100.56     3600.11     4601.01
2026-07-02 12:00:00   6800.23     6100.78     7499.68  ← 大促次日

# 基于 Prophet 预测的 HPA 策略
apiVersion:autoscaling/v2
kind:HorizontalPodAutoscaler
metadata:
name:order-service-hpa
namespace:production
spec:
scaleTargetRef:
apiVersion:apps/v1
kind:Deployment
name:order-service
minReplicas:3
maxReplicas:20
behavior:
scaleUp:
policies:
-type:Pods
value:4# 每次最多扩 4 个 Pod
periodSeconds:60
stabilizationWindowSeconds:0# 预扩容不等待
scaleDown:
stabilizationWindowSeconds:300# 缩容等 5 分钟确认
metrics:
-type:Pods
pods:
metric:
name:http_requests_per_second
target:
type:AverageValue
averageValue:"500"# 单 Pod 承载 500 QPS

"""
容量基线计算 + 扩缩容决策（关键函数）
"""

import math

defcalc_required_pods(
    predicted_qps: float,
    qps_per_pod: int = 500,
    safety_margin: float = 0.8# 80% 水位线触发扩容
) -> dict:
"""
    根据预测 QPS 计算所需 Pod 数
    safety_margin: 保留 20% 的余量应对突发
    """
    target_qps = predicted_qps / safety_margin
    required = math.ceil(target_qps / qps_per_pod)

return {
"predicted_qps": round(predicted_qps, 1),
"target_qps_with_margin": round(target_qps, 1),
"required_pods": required,
"safety_margin": f"{safety_margin*100}%"
    }

# 示例：大促日峰值预测 6800 QPS
result = calc_required_pods(predicted_qps=6800, qps_per_pod=500)
print(f"📈 预测峰值: {result['predicted_qps']} QPS")
print(f"🛡️ 含安全裕度: {result['target_qps_with_margin']} QPS")
print(f"📦 所需 Pod 数: {result['required_pods']} 个")

📈 预测峰值: 6800.0 QPS
🛡️ 含安全裕度: 8500.0 QPS
📦 所需 Pod 数: 17 个

from statsmodels.tsa.stattools import adfuller

defcheck_stationarity(series):
"""ADF 检验，判断序列是否平稳"""
    result = adfuller(series.dropna())
    p_value = result[1]

if p_value < 0.05:
print(f"✅ 序列平稳，d=0")
return0
else:
        diff1 = series.diff().dropna()
        result1 = adfuller(diff1)
if result1[1] < 0.05:
print(f"✅ 一阶差分后平稳，d=1")
return1
print("❌ 建议换用 Prophet")
return2

# 关键配置
scaleDown:
stabilizationWindowSeconds:300# 缩容等 5 分钟确认

#!/bin/bash
# =====================================================
# capacity_forecast_cron.sh — AI 容量预测自动化
# =====================================================

PROMETHEUS="http://localhost:9090"
MODEL_DIR="/opt/models/capacity"
LOG_DIR="/var/log/capacity_forecast"
DATE=$(date +%Y%m%d)

mkdir -p $LOG_DIR

# Step 1: 拉取最新训练数据（365 天滚动）
python3 /opt/scripts/fetch_training_data.py \
    --prometheus $PROMETHEUS \
    --days 365 \
    --output $MODEL_DIR/training_data.csv

# Step 2: Prophet 预测未来 7 天
python3 /opt/scripts/run_forecast.py \
    --data $MODEL_DIR/training_data.csv \
    --period 168 --freq h \
    --output $MODEL_DIR/forecast_result.json

# Step 3: 检查业务日历（今天是否大促？）
python3 /opt/scripts/check_business_calendar.py \
    --forecast $MODEL_DIR/forecast_result.json

# Step 4: 更新 HPA 容量基线
python3 /opt/scripts/update_hpa.py \
    --forecast $MODEL_DIR/forecast_result.json \
    --kubeconfig /etc/kubernetes/admin.conf

# Step 5: 回测验证（保证精度）
python3 /opt/scripts/backtest.py \
    --data $MODEL_DIR/training_data.csv \
    --threshold 0.15 \
    >> $LOG_DIR/backtest_history.csv

echo"[$(date)] 容量预测完成 ✅" | tee -a $LOG_DIR/$DATE.log

# Crontab 配置
# 每天凌晨 3 点预测
0 3 * * * /opt/scripts/capacity_forecast_cron.sh

# 每 6 小时检查预警（大促前加为每小时）
0 */6 * * * /opt/scripts/check_alert.sh

# 大促前三天加频
0 0 15,16,17,18 6 * /opt/scripts/pre_da_cu.sh