recommend-emr-dataprocess/ad/02_ad_model_update_twice_daily.sh

#!/bin/sh
set -x


export SPARK_HOME=/opt/apps/SPARK2/spark-2.4.8-hadoop3.2-1.0.8
export PATH=$SPARK_HOME/bin:$PATH
export HADOOP_CONF_DIR=/etc/taihao-apps/hadoop-conf
export JAVA_HOME=/usr/lib/jvm/java-1.8.0

source /root/anaconda3/bin/activate py37

# 全局常量
originDataSavePath=/dw/recommend/model/31_ad_sample_data_v4_auto
bucketFeatureSavePathHome=/dw/recommend/model/33_ad_train_data_v4_auto
model_name=model_bkb8_v4
LAST_MODEL_HOME=/root/zhaohp/model_online

MODEL_HOME=/root/zhaohp/recommend-emr-dataprocess/model
OSS_PATH=oss://art-recommend.oss-cn-hangzhou.aliyuncs.com/ad_model

PREDICT_PATH=/root/zhaohp/recommend-emr-dataprocess/predict
HADOOP=/opt/apps/HADOOP-COMMON/hadoop-common-current/bin/hadoop
FM_HOME=/root/sunmingze/alphaFM

today="$(date +%Y%m%d)"
today_early_1="$(date -d '1 days ago' +%Y%m%d)"

start_time=$(date +%s)
elapsed=0
LOG_PREFIX=广告模型自动更新任务

# 训练和预测数据分区
train_begin_str=''
train_end_str=''
predict_begin_str=''
predict_end_str=''

# HDFS保存数据的目录
trainBucketFeaturePath=${bucketFeatureSavePathHome}
predictBucketFeaturePath=${bucketFeatureSavePathHome}

local_model_file_path=${MODEL_HOME}/${model_name}.txt
local_change_model_file_path=${MODEL_HOME}/${model_name}_change.txt

max_hour=21
max_minute=20

# 全局初始化
global_init() {
    # 获取当前小时，确定需要使用的数据分区范围
    local current_hour="$(date +%H)"
    if [ $current_hour -le 06 ]; then
        train_begin_str=${today_early_1}08
        train_end_str=${today_early_1}21
        predict_begin_str=${today_early_1}22
        predict_end_str=${today_early_1}23

        trainBucketFeaturePath=${bucketFeatureSavePathHome}/${today_early_1}/train
        predictBucketFeaturePath=${bucketFeatureSavePathHome}/${today_early_1}/predict

        local_model_file_path=${MODEL_HOME}/${model_name}_${train_end_str}.txt
        local_change_model_file_path=${MODEL_HOME}/${model_name}_${train_end_str}_change.txt
        max_hour=08
    elif [ $current_hour -ge 16 ]; then 
        train_begin_str=${today_early_1}22
        train_end_str=${today}13
        predict_begin_str=${today}14
        predict_end_str=${today}15

        trainBucketFeaturePath=${bucketFeatureSavePathHome}/${today}/train
        predictBucketFeaturePath=${bucketFeatureSavePathHome}/${today}/predict

        local_model_file_path=${MODEL_HOME}/${model_name}_${train_end_str}.txt
        local_change_model_file_path=${MODEL_HOME}/${model_name}_${train_end_str}_change.txt
        max_hour=21

    else
        echo "当前时间段异常: 退出任务"
        exit 1
    fi

    # 删除HDFS目录，保证本次任务运行时目录干净
    $HADOOP fs -rm -r -skipTrash ${trainBucketFeaturePath}
    $HADOOP fs -rm -r -skipTrash ${predictBucketFeaturePath}

    echo "全局变量初始化化: "
    echo "  train_begin_str=${train_begin_str}"
    echo "  train_end_str=${train_end_str}"
    echo "  predict_begin_str=${predict_begin_str}"
    echo "  predict_end_str=${predict_end_str}"
    echo "  originDataSavePath=${originDataSavePath}"
    echo "  trainBucketFeaturePath=${trainBucketFeaturePath}"
    echo "  predictBucketFeaturePath=${predictBucketFeaturePath}"
    echo "  local_model_file_path=${local_model_file_path}"
    echo "  local_change_model_file_path=${local_change_model_file_path}"
    echo "  max_hour=${max_hour}"

}

# 校验命令的退出码
check_run_status() {
    local status=$1
    local step_start_time=$2
    local step_name=$3

    local step_end_time=$(date +%s)
    local step_elapsed=$(($step_end_time - $step_start_time))

    if [ $status -ne 0 ]; then
        echo "$LOG_PREFIX -- ${step_name}失败: 耗时 $step_elapsed"
        local elapsed=$(($step_end_time - $start_time))
        # /root/anaconda3/bin/python ad/ad_monitor_util.py --level error --msg "$msg" --start "$start_time" --elapsed "$elapsed"
        exit 1
    else
        echo "$LOG_PREFIX -- ${step_name}成功: 耗时 $step_elapsed"
    fi
}

# 校验大数据任务是否执行完成
check_ad_hive() {
    local step_start_time=$(date +%s)
    while true; do
        local python_return_code=$(python ad/ad_utils.py --excute_program check_ad_origin_hive --partition ${predict_end_str:0:8} --hh ${predict_end_str:8:10})

        local step_end_time=$(date +%s)
        local elapsed=$(($step_end_time - $step_start_time))
        if [ "$python_return_code" -eq 0 ]; then
            break
        fi
        echo "Python程序返回非0值，等待五分钟后再次调用。"
        sleep 300
        local current_hour=$(date +%H)
        local current_minute=$(date +%M)
        if (( current_hour > max_hour || (current_hour == max_hour && current_minute >= max_minute) )); then
            local msg="大数据数据生产校验失败, 分区: ${today}10"
            echo -e "$LOG_PREFIX -- 大数据数据生产校验 -- ${msg}: 耗时 $elapsed"
            # /root/anaconda3/bin/python ad/ad_monitor_util.py --level error --msg "$msg" --start "$start_time" --elapsed "$elapsed"
            exit 1
        fi
    done
    echo "$LOG_PREFIX -- 大数据数据生产校验 -- 大数据数据生产校验通过: 耗时 $elapsed"

}

# 原始特征生产
make_origin_data() {
    local step_start_time=$(date +%s)
    /opt/apps/SPARK2/spark-2.4.8-hadoop3.2-1.0.8/bin/spark-class2 org.apache.spark.deploy.SparkSubmit \
    --class com.aliyun.odps.spark.zhp.makedata_ad.makedata_ad_31_originData_20240620 \
    --master yarn --driver-memory 1G --executor-memory 2G --executor-cores 1 --num-executors 16 \
    ./target/spark-examples-1.0.0-SNAPSHOT-shaded.jar \
    tablePart:64 repartition:16 \
    beginStr:${train_begin_str} endStr:${predict_end_str} \
    savePath:${originDataSavePath} \
    table:alg_recsys_ad_sample_all filterHours:00,01,02,03,04,05,06,07 \
    idDefaultValue:0.01

    local return_code=$?
    check_run_status $return_code $step_start_time "Spark原始样本生产任务"

}

# 训练用数据分桶
make_train_bucket_feature() {
    /opt/apps/SPARK2/spark-2.4.8-hadoop3.2-1.0.8/bin/spark-class2 org.apache.spark.deploy.SparkSubmit \
    --class com.aliyun.odps.spark.zhp.makedata_ad.makedata_ad_33_bucketData_20240717 \
    --master yarn --driver-memory 2G --executor-memory 4G --executor-cores 1 --num-executors 16 \
    ./target/spark-examples-1.0.0-SNAPSHOT-shaded.jar \
    beginStr:${train_begin_str:0:8} endStr:${train_end_str:0:8} repartition:100 \
    filterNames:adid_,targeting_conversion_ \
    readPath:${originDataSavePath} \
    savePath:${trainBucketFeaturePath}
}

# 预测用数据分桶
make_predict_bucket_feature() {
    /opt/apps/SPARK2/spark-2.4.8-hadoop3.2-1.0.8/bin/spark-class2 org.apache.spark.deploy.SparkSubmit \
    --class com.aliyun.odps.spark.zhp.makedata_ad.makedata_ad_33_bucketData_20240717 \
    --master yarn --driver-memory 2G --executor-memory 4G --executor-cores 1 --num-executors 16 \
    ./target/spark-examples-1.0.0-SNAPSHOT-shaded.jar \
    beginStr:${predict_begin_str:0:8} endStr:${predict_end_str:0:8} repartition:100 \
    filterNames:adid_,targeting_conversion_ \
    readPath:${originDataSavePath} \
    savePath:${predictBucketFeaturePath}
}


# 特征分桶，训练用的数据和预测用的数据分不同的目录
make_bucket_feature() {
    local step_start_time=$(date +%s)
    
    # 训练用的数据
    make_train_bucket_feature &
    train_bucket_pid=$!

    wait $train_bucket_pid

    local train_return_code=$?
    check_run_status $train_return_code $step_start_time "Spark特征分桶任务: 训练数据分桶"

    
    # 预测用的数据
    make_predict_bucket_feature &
    predict_bucket_pid=$!

    wait $predict_bucket_pid

    local predict_return_code=$?
    check_run_status $predict_return_code $step_start_time "Spark特征分桶任务: 预测数据分桶"
}

# 模型训练
model_train() {
    local step_start_time=$(date +%s)
    $HADOOP fs -text ${trainBucketFeaturePath}/*/* | ${FM_HOME}/bin/fm_train -m ${local_model_file_path} -dim 1,1,8  -im ${LAST_MODEL_HOME}/model_online.txt -core 8

    local return_code=$?
    check_run_status $return_code $step_start_time "模型训练"
}


# 计算线上模型的AUC
calc_online_model_auc() {
    $HADOOP fs -text ${predictBucketFeaturePath}/*/* | ${FM_HOME}/bin/fm_predict -m ${LAST_MODEL_HOME}/model_online.txt -dim 8 -core 8 -out ${PREDICT_PATH}/${model_name}_${train_end_str}_online.txt
    online_auc=`cat ${PREDICT_PATH}/${model_name}_${train_end_str}_online.txt | /root/sunmingze/AUC/AUC`
}


# 计算新模型AUC
calc_new_model_auc() {
    $HADOOP fs -text ${predictBucketFeaturePath}/*/* | ${FM_HOME}/bin/fm_predict -m ${local_model_file_path} -dim 8 -core 8 -out ${PREDICT_PATH}/${model_name}_${train_end_str}_new.txt
    new_auc=`cat ${PREDICT_PATH}/${model_name}_${train_end_str}_new.txt | /root/sunmingze/AUC/AUC`
}

# AUC对比
auc_compare() {
    local step5_start_time=$(date +%s)

    # 5.1 计算线上模型的AUC
    local step_start_time=$(date +%s)

    calc_online_model_auc &
    local calc_online_model_auc_pid=$!

    wait $calc_online_model_auc_pid
    local return_code=$?
    check_run_status $return_code $step_start_time "线上模型AUC计算"

    # 5.2 计算新模型的AUC
    step_start_time=$(date +%s)

    calc_new_model_auc &
    local calc_new_model_auc_pid=$!

    wait $calc_new_model_auc_pid

    local new_return_code=$?
    check_run_status $new_return_code $step_start_time "新模型的AUC计算"



    echo "AUC比对: 线上模型的AUC: ${online_auc}, 新模型的AUC: ${new_auc}"

    # 5.3 计算新模型与线上模型的AUC差值的绝对值
    auc_diff=$(echo "$online_auc - $new_auc" | bc -l)
    local auc_diff_abs=$(echo "sqrt(($auc_diff)^2)" | bc -l)

    local step_end_time=$(date +%s)
    local step5_elapsed=$(($step_end_time - $step5_start_time))
    # 5.4 如果差值的绝对值小于0.005且新模型的AUC大于0.73, 则更新模型
    if (( $(echo "${online_auc} <= ${new_auc}" | bc -l) )); then
        local msg="新模型优于线上模型 \n\t线上模型AUC: ${online_auc} \n\t新模型AUC: ${new_auc}"
        echo -e "$LOG_PREFIX -- $msg: 耗时 $step5_elapsed"

    elif (( $(echo "$auc_diff_abs < 0.005" | bc -l) )) && (( $(echo "$new_auc >= 0.73" | bc -l) )); then
        local msg="新模型与线上模型差值小于阈值0.005 \n\t线上模型AUC: ${online_auc} \n\t新模型AUC: ${new_auc} \n\t差值为: $auc_diff_abs"
        echo -e "$LOG_PREFIX -- $msg: 耗时 $step5_elapsed"

    else
        local msg="新模型与线上模型差值大于等于阈值0.005或新模型的AUC小于0.73 \n\t线上模型AUC: ${online_auc} \n\t新模型AUC: ${new_auc} \n\t差值为: $auc_diff"
        echo -e "$LOG_PREFIX -- $msg: 耗时 $step5_elapsed"
        local elapsed=$(($step_end_time - $start_time))
        # /root/anaconda3/bin/python ad/ad_monitor_util.py --level error --msg "$msg" --start "$start_time" --elapsed "$elapsed"
        exit 1
    fi
}

# 模型格式转换
model_to_online_format() {
    local step_start_time=$(date +%s)
    cat ${local_model_file_path} |
    awk -F " " '{
        if (NR == 1) {
            print $1"\t"$2
        } else {
            split($0, fields, " ");
            OFS="\t";
            line=""
            for (i = 1; i <= 10 && i <= length(fields); i++) {
                line = (line ? line "\t" : "") fields[i];
            }
            print line
        }
    }' > ${local_change_model_file_path}

    local return_code=$?
    check_run_status $return_code $step_start_time "模型格式转换"

}

# 模型文件上传OSS
model_upload_oss() {
    local step_start_time=$(date +%s)
    local online_model_path=${OSS_PATH}/${model_name}.txt
    $HADOOP fs -test -e ${online_model_path}
    if [ $? -eq 0 ]; then
        echo "删除已存在的OSS模型文件"
        $HADOOP fs -rm -r -skipTrash ${online_model_path}
    fi
    $HADOOP fs -put ${local_change_model_file_path} ${online_model_path}
    
    local return_code=$?
    check_run_status $return_code $step_start_time "模型文件上传OSS"
    
}

# 模型文件本地备份
model_local_back() {
    local step_start_time=$(date +%s)
    # 将之前的线上模型进行备份，表示从上一个备份时间到当前时间内，使用的线上模型都是此文件
    # 假设当前是07-11，上一次备份时间为07-07。备份之后表示从07-07下午至07-11上午线上使用的模型文件都是model_online_20240711.txt
    file_suffix=$(date "+%Y%m%d%H")
    cp -f ${LAST_MODEL_HOME}/model_online.txt ${LAST_MODEL_HOME}/model_online_${file_suffix}.txt
    cp -f ${local_model_file_path} ${LAST_MODEL_HOME}/model_online.txt

    local return_code=$?
    check_run_status $return_code $step_start_time "模型备份"
}

# 任务完成通知
success_inform() {
    local step_end_time=$(date +%s)
    local msg="\n\t - 广告模型文件更新完成 \n\t - 前一天线上模型全天数据AUC: $yesterday_online_auc \n\t - 前一天新模型全天数据AUC: $yesterday_new_auc \n\t - 新模型AUC: $new_auc \n\t - 线上模型AUC: $online_auc  \n\t - AUC差值: $auc_diff \n\t - 模型上传路径: $online_model_path"
    echo -e "$LOG_PREFIX -- 模型更新完成 -- $msg: 耗时 $step_elapsed"
    local elapsed=$(($step_end_time - $start_time))
    # /root/anaconda3/bin/python ad/ad_monitor_util.py --level info --msg "$msg" --start "$start_time" --elapsed "$elapsed"
}

main() {

    global_init

    check_ad_hive

    make_origin_data

    make_bucket_feature

    # model_train

    # auc_compare

    # model_to_online_format

    # model_upload_oss

    # model_local_back

    # success_inform
}


main



# nohup ./ad/02_ad_model_update_twice_daily.sh > logs/02_twice_daily.log 2>&1 &