rov-offline/alg_recsys_recall_24h_region.py

# -*- coding: utf-8 -*-
import multiprocessing
import traceback
import gevent
import datetime
import pandas as pd
import math
from functools import reduce
from odps import ODPS
from threading import Timer, Thread
from my_utils import RedisHelper, get_data_from_odps, filter_video_status, check_table_partition_exits, \
    filter_video_status_app, send_msg_to_feishu
from my_config import set_config
from log import Log

# os.environ['NUMEXPR_MAX_THREADS'] = '16'

config_, _ = set_config()
log_ = Log()

region_code = config_.REGION_CODE


RULE_PARAMS = {
    'rule_params': {
        'rule66': {'view_type': 'video-show', 'return_count': 21, 'score_rule': 0, 'platform_return_rate': 0.001},
    },
    'data_params': config_.DATA_PARAMS,
    'params_list': [
        {'data': 'data66', 'rule': 'rule66'},
    ]
}

features = [
    'apptype',
    'code',  # 省份编码
    'videoid',
    'lastday_preview',  # 昨日预曝光人数
    'lastday_view',  # 昨日曝光人数
    'lastday_play',  # 昨日播放人数
    'lastday_share',  # 昨日分享人数
    'lastday_return',  # 昨日回流人数
    'lastday_preview_total',  # 昨日预曝光次数
    'lastday_view_total',  # 昨日曝光次数
    'lastday_play_total',  # 昨日播放次数
    'lastday_share_total',  # 昨日分享次数
    'platform_return',
    'platform_preview',
    'platform_preview_total',
    'platform_show',
    'platform_show_total',
    'platform_view',
    'platform_view_total',
]


def get_rov_redis_key(now_date):
    """获取rov模型结果存放key"""
    redis_helper = RedisHelper()
    now_dt = datetime.datetime.strftime(now_date, '%Y%m%d')
    key_name = f'{config_.RECALL_KEY_NAME_PREFIX}{now_dt}'
    if not redis_helper.key_exists(key_name=key_name):
        pre_dt = datetime.datetime.strftime(now_date - datetime.timedelta(days=1), '%Y%m%d')
        key_name = f'{config_.RECALL_KEY_NAME_PREFIX}{pre_dt}'
    return key_name


def data_check(project, table, now_date):
    """检查数据是否准备好"""
    odps = ODPS(
        access_id=config_.ODPS_CONFIG['ACCESSID'],
        secret_access_key=config_.ODPS_CONFIG['ACCESSKEY'],
        project=project,
        endpoint=config_.ODPS_CONFIG['ENDPOINT'],
        connect_timeout=3000,
        read_timeout=500000,
        pool_maxsize=1000,
        pool_connections=1000
    )

    try:
        dt = datetime.datetime.strftime(now_date, '%Y%m%d%H')
        check_res = check_table_partition_exits(date=dt, project=project, table=table)
        if check_res:
            sql = f'select * from {project}.{table} where dt = {dt}'
            with odps.execute_sql(sql=sql).open_reader() as reader:
                data_count = reader.count
        else:
            data_count = 0
    except Exception as e:
        data_count = 0
    return data_count


def get_feature_data(project, table, now_date):
    """获取特征数据"""
    dt = datetime.datetime.strftime(now_date, '%Y%m%d%H')
    # dt = '2022041310'
    records = get_data_from_odps(date=dt, project=project, table=table)
    feature_data = []
    for record in records:
        item = {}
        for feature_name in features:
            item[feature_name] = record[feature_name]
        feature_data.append(item)
    feature_df = pd.DataFrame(feature_data)
    return feature_df


def cal_score(df, param):
    """
    计算score
    :param df: 特征数据
    :param param:
    :return:
    """
    # score计算公式: sharerate*backrate*logback*ctr
    # sharerate = lastday_share/(lastday_play+1000)
    # backrate = lastday_return/(lastday_share+10)
    # ctr = lastday_play/(lastday_preview+1000), 对ctr限最大值：K2 = 0.6 if ctr > 0.6 else ctr
    # score = sharerate * backrate * LOG(lastday_return+1) * K2

    df = df.fillna(0)
    df['share_rate'] = df['lastday_share'] / (df['lastday_play'] + 1000)
    df['back_rate'] = df['lastday_return'] / (df['lastday_share'] + 10)
    df['log_back'] = (df['lastday_return'] + 1).apply(math.log)
    if param.get('view_type', None) == 'video-show':
        df['ctr'] = df['lastday_play'] / (df['platform_show'] + 1000)
    else:
        df['ctr'] = df['lastday_play'] / (df['lastday_preview'] + 1000)
    df['K2'] = df['ctr'].apply(lambda x: 0.6 if x > 0.6 else x)
    df['platform_return_rate'] = df['platform_return'] / df['lastday_return']

    df['score1'] = df['share_rate'] * df['back_rate'] * df['log_back'] * df['K2']

    click_score_rate = param.get('click_score_rate', None)
    back_score_rate = param.get('click_score_rate', None)
    if click_score_rate is not None:
        df['score'] = (1 - click_score_rate) * df['score1'] + click_score_rate * df['K2']
    elif back_score_rate is not None:
        df['score'] = (1 - back_score_rate) * df['score1'] + back_score_rate * df['back_rate']
    else:
        df['score'] = df['score1']

    df = df.sort_values(by=['score'], ascending=False)
    return df


def video_rank(df, now_date, now_h, rule_key, param, region, data_key):
    """
    获取符合进入召回源条件的视频
    :param df:
    :param now_date:
    :param now_h:
    :param rule_key: 小时级数据进入条件
    :param param: 小时级数据进入条件参数
    :param region: 所属地域
    :return:
    """
    redis_helper = RedisHelper()
    # 获取符合进入召回源条件的视频
    return_count = param.get('return_count', 1)
    score_value = param.get('score_rule', 0)
    platform_return_rate = param.get('platform_return_rate', 0)
    h_recall_df = df[(df['lastday_return'] >= return_count) & (df['score'] >= score_value)
                     & (df['platform_return_rate'] >= platform_return_rate)]
    # videoid重复时，保留分值高
    h_recall_df = h_recall_df.sort_values(by=['score'], ascending=False)
    h_recall_df = h_recall_df.drop_duplicates(subset=['videoid'], keep='first')
    h_recall_df['videoid'] = h_recall_df['videoid'].astype(int)
    h_recall_videos = h_recall_df['videoid'].to_list()
    log_.info(f"各种规则过滤后，一共有多少个视频 = {len(h_recall_videos)}")
    # 视频状态过滤
    if data_key in ['data7', ]:
        filtered_videos = filter_video_status_app(h_recall_videos)
    else:
        filtered_videos = filter_video_status(h_recall_videos)
    log_.info(f"视频状态-过滤后，一共有多少个视频 = {len(filtered_videos)}")

    # 写入对应的redis
    h_video_ids = []
    day_recall_result = {}
    for video_id in filtered_videos:
        score = h_recall_df[h_recall_df['videoid'] == video_id]['score']
        day_recall_result[int(video_id)] = float(score)
        h_video_ids.append(int(video_id))
    day_recall_key_name = \
        f"{config_.RECALL_KEY_NAME_PREFIX_REGION_BY_24H}{region}:{data_key}:{rule_key}:" \
        f"{datetime.datetime.strftime(now_date, '%Y%m%d')}:{now_h}"
    log_.info("打印地域24小时的某个地域{},redis key:{}".format(region, day_recall_key_name))
    if len(day_recall_result) > 0:
        log_.info(f"开始写入头部数据：count = {len(day_recall_result)}, key = {day_recall_key_name}")
        redis_helper.add_data_with_zset(key_name=day_recall_key_name, data=day_recall_result, expire_time=2 * 3600)
    else:
        log_.info(f"无数据，不写入。")
        # 清空线上过滤应用列表
        # redis_helper.del_keys(key_name=f"{config_.REGION_H_VIDEO_FILER_24H}{region}.{app_type}.{data_key}.{rule_key}")

    # 与其他召回视频池去重，存入对应的redis
    # dup_to_redis(h_video_ids=h_video_ids, now_date=now_date, now_h=now_h, rule_key=rule_key, region=region)


def merge_df(df_left, df_right):
    """
    df按照videoid, code 合并，对应特征求和
    :param df_left:
    :param df_right:
    :return:
    """
    df_merged = pd.merge(df_left, df_right, on=['videoid', 'code'], how='outer', suffixes=['_x', '_y'])
    df_merged.fillna(0, inplace=True)
    feature_list = ['videoid', 'code']
    for feature in features:
        if feature in ['apptype', 'videoid', 'code']:
            continue
        df_merged[feature] = df_merged[f'{feature}_x'] + df_merged[f'{feature}_y']
        feature_list.append(feature)
    return df_merged[feature_list]


def merge_df_with_score(df_left, df_right):
    """
    df 按照[videoid, code]合并，平台回流人数、回流人数、分数 分别求和
    :param df_left:
    :param df_right:
    :return:
    """
    df_merged = pd.merge(df_left, df_right, on=['videoid', 'code'], how='outer', suffixes=['_x', '_y'])
    df_merged.fillna(0, inplace=True)
    feature_list = ['videoid', 'code', 'lastday_return', 'platform_return', 'score']
    for feature in feature_list[2:]:
        df_merged[feature] = df_merged[f'{feature}_x'] + df_merged[f'{feature}_y']
    return df_merged[feature_list]


def process_with_region(region, df_merged, data_key, rule_key, rule_param, now_date, now_h):
    log_.info(f"多协程的region = {region} 开始执行")
    region_df = df_merged[df_merged['code'] == region]
    log_.info(f'该区域region = {region}, 下有多少数据量 = {len(region_df)}')
    score_df = cal_score(df=region_df, param=rule_param)
    video_rank(df=score_df, now_date=now_date, now_h=now_h, region=region,
               rule_key=rule_key, param=rule_param, data_key=data_key)
    log_.info(f"多协程的region = {region} 完成执行")


def process_with_region2(region, df_merged, data_key, rule_key, rule_param, now_date, now_h):
    log_.info(f"region = {region} start...")
    region_score_df = df_merged[df_merged['code'] == region]
    log_.info(f'region = {region}, region_score_df count = {len(region_score_df)}')
    video_rank(df=region_score_df, now_date=now_date, now_h=now_h, region=region,
               rule_key=rule_key, param=rule_param, data_key=data_key)
    log_.info(f"region = {region} end!")


def process_with_app_type(app_type, params, region_code_list, feature_df, now_date, now_h):
    log_.info(f"app_type = {app_type} start...")
    data_params_item = params.get('data_params')
    rule_params_item = params.get('rule_params')
    for param in params.get('params_list'):
        data_key = param.get('data')
        data_param = data_params_item.get(data_key)
        log_.info(f"data_key = {data_key}, data_param = {data_param}")
        df_list = [feature_df[feature_df['apptype'] == apptype] for apptype in data_param]
        df_merged = reduce(merge_df, df_list)

        rule_key = param.get('rule')
        rule_param = rule_params_item.get(rule_key)
        log_.info(f"rule_key = {rule_key}, rule_param = {rule_param}")
        task_list = [
            gevent.spawn(process_with_region, region, df_merged, app_type, data_key, rule_key, rule_param,
                         now_date, now_h)
            for region in region_code_list
        ]
        gevent.joinall(task_list)
    log_.info(f"app_type = {app_type} end!")


def process_with_param(param, data_params_item, rule_params_item, region_code_list, feature_df, now_date, now_h):
    data_key = param.get('data')
    data_param = data_params_item.get(data_key)
    rule_key = param.get('rule')
    rule_param = rule_params_item.get(rule_key)
    merge_func = rule_param.get('merge_func', None)
    log_.info("数据采用:{},统计采用{}.".format(data_key, rule_key))
    log_.info("具体的规则是:{}.".format(rule_param))

    if merge_func == 2:
        score_df_list = []
        for apptype, weight in data_param.items():
            df = feature_df[feature_df['apptype'] == apptype]
            # 计算score
            score_df = cal_score(df=df, param=rule_param)
            score_df['score'] = score_df['score'] * weight
            score_df_list.append(score_df)
        # 分数合并
        df_merged = reduce(merge_df_with_score, score_df_list)
        # 更新平台回流比
        df_merged['platform_return_rate'] = df_merged['platform_return'] / df_merged['lastday_return']
        task_list = [
            gevent.spawn(process_with_region2, region, df_merged, data_key, rule_key, rule_param, now_date, now_h)
            for region in region_code_list
        ]
    else:
        df_list = [feature_df[feature_df['apptype'] == apptype] for apptype, _ in data_param.items()]
        df_merged = reduce(merge_df, df_list)
        task_list = [
            gevent.spawn(process_with_region, region, df_merged, data_key, rule_key, rule_param, now_date, now_h)
            for region in region_code_list
        ]

    gevent.joinall(task_list)
    log_.info(f"多进程的 param = {param} 完成执行!")


def rank_by_24h(project, table, now_date, now_h, rule_params, region_code_list):
    # 获取特征数据
    feature_df = get_feature_data(project=project, table=table, now_date=now_date)
    feature_df['apptype'] = feature_df['apptype'].astype(int)
    # rank
    data_params_item = rule_params.get('data_params')
    rule_params_item = rule_params.get('rule_params')
    params_list = rule_params.get('params_list')
    pool = multiprocessing.Pool(processes=len(params_list))
    for param in params_list:
        pool.apply_async(
            func=process_with_param,
            args=(param, data_params_item, rule_params_item, region_code_list, feature_df, now_date, now_h)
        )
    pool.close()
    pool.join()

    """
    pool = multiprocessing.Pool(processes=len(config_.APP_TYPE))
    for app_type, params in rule_params.items():
        pool.apply_async(func=process_with_app_type,
                         args=(app_type, params, region_code_list, feature_df, now_date, now_h))
    pool.close()
    pool.join()
    """


def dup_to_redis(h_video_ids, now_date, now_h, rule_key, region):
    """将地域分组小时级数据与其他召回视频池去重，存入对应的redis"""
    redis_helper = RedisHelper()

    # ##### 去重小程序天级更新结果，并另存为redis中
    day_key_name = f"{config_.RECALL_KEY_NAME_PREFIX_BY_DAY}rule2.{datetime.datetime.strftime(now_date, '%Y%m%d')}"
    if redis_helper.key_exists(key_name=day_key_name):
        day_data = redis_helper.get_all_data_from_zset(key_name=day_key_name, with_scores=True)
        log_.info(f'day data count = {len(day_data)}')
        day_dup = {}
        for video_id, score in day_data:
            if int(video_id) not in h_video_ids:
                day_dup[int(video_id)] = score
                h_video_ids.append(int(video_id))
        log_.info(f"day data dup count = {len(day_dup)}")
        day_dup_key_name = \
            f"{config_.RECALL_KEY_NAME_PREFIX_DUP_REGION_DAY_24H}{region}.{rule_key}." \
            f"{datetime.datetime.strftime(now_date, '%Y%m%d')}.{now_h}"
        if len(day_dup) > 0:
            redis_helper.add_data_with_zset(key_name=day_dup_key_name, data=day_dup, expire_time=23 * 3600)

    # ##### 去重小程序模型更新结果，并另存为redis中
    model_key_name = get_rov_redis_key(now_date=now_date)
    model_data = redis_helper.get_all_data_from_zset(key_name=model_key_name, with_scores=True)
    log_.info(f'model data count = {len(model_data)}')
    model_data_dup = {}
    for video_id, score in model_data:
        if int(video_id) not in h_video_ids:
            model_data_dup[int(video_id)] = score
            h_video_ids.append(int(video_id))
    log_.info(f"model data dup count = {len(model_data_dup)}")
    model_data_dup_key_name = \
        f"{config_.RECALL_KEY_NAME_PREFIX_DUP_REGION_24H}{region}.{rule_key}." \
        f"{datetime.datetime.strftime(now_date, '%Y%m%d')}.{now_h}"
    if len(model_data_dup) > 0:
        redis_helper.add_data_with_zset(key_name=model_data_dup_key_name, data=model_data_dup, expire_time=23 * 3600)


def h_rank_bottom(now_date, now_h, rule_params, region_code_list):
    """未按时更新数据，用上一小时结果作为当前小时的数据"""
    redis_helper = RedisHelper()
    if now_h == 0:
        redis_dt = datetime.datetime.strftime(now_date - datetime.timedelta(days=1), '%Y%m%d')
        redis_h = 23
    else:
        redis_dt = datetime.datetime.strftime(now_date, '%Y%m%d')
        redis_h = now_h - 1

    # 以上一小时的地域分组数据作为当前小时的数据
    key_prefix = config_.RECALL_KEY_NAME_PREFIX_REGION_BY_24H
    for param in rule_params.get('params_list'):
        data_key = param.get('data')
        rule_key = param.get('rule')
        log_.info(f"data_key = {data_key}, rule_key = {rule_key}")
        for region in region_code_list:
            log_.info(f"region = {region}")
            key_name = f"{key_prefix}{region}:{data_key}:{rule_key}:{redis_dt}:{redis_h}"
            initial_data = redis_helper.get_all_data_from_zset(key_name=key_name, with_scores=True)
            if initial_data is None:
                initial_data = []
            final_data = dict()
            h_video_ids = []
            for video_id, score in initial_data:
                final_data[video_id] = score
                h_video_ids.append(int(video_id))
            # 存入对应的redis
            final_key_name = \
                f"{key_prefix}{region}:{data_key}:{rule_key}:{datetime.datetime.strftime(now_date, '%Y%m%d')}:{now_h}"
            if len(final_data) > 0:
                redis_helper.add_data_with_zset(key_name=final_key_name, data=final_data, expire_time=2 * 3600)




def h_timer_check():
    try:
        rule_params = RULE_PARAMS
        project = config_.PROJECT_REGION_24H_APP_TYPE
        table = config_.TABLE_REGION_24H_APP_TYPE
        region_code_list = [code for region, code in region_code.items() if code != '-1']
        now_date = datetime.datetime.today()
        now_h = datetime.datetime.now().hour
        now_min = datetime.datetime.now().minute
        log_.info(f"开始执行: {datetime.datetime.strftime(now_date, '%Y%m%d%H')}")
        # 查看当天更新的数据是否已准备好
        h_data_count = data_check(project=project, table=table, now_date=now_date)
        if h_data_count > 0:
            log_.info('上游数据表查询数据条数 h_data_count = {}，开始计算。'.format(h_data_count))
            rank_by_24h(now_date=now_date, now_h=now_h, rule_params=rule_params,
                        project=project, table=table, region_code_list=region_code_list)
            log_.info("数据3----------正常完成----------")
        elif now_min > 40:
            log_.info('当前分钟超过40，预计执行无法完成，使用 bottom data!')
            h_rank_bottom(now_date=now_date, now_h=now_h, rule_params=rule_params, region_code_list=region_code_list)
            log_.info('----------当前分钟超过40，使用bottom的data，完成----------')
        else:
            # 数据没准备好，1分钟后重新检查
            log_.info("上游数据未就绪，等待...")
            Timer(60, h_timer_check).start()

    except Exception as e:
        log_.error(f"地域分组24h数据更新失败, exception: {e}, traceback: {traceback.format_exc()}")
        send_msg_to_feishu(
            webhook=config_.FEISHU_ROBOT['server_robot'].get('webhook'),
            key_word=config_.FEISHU_ROBOT['server_robot'].get('key_word'),
            msg_text=f"rov-offline{config_.ENV_TEXT} - 地域分组24h数据更新失败\n"
                     f"exception: {e}\n"
                     f"traceback: {traceback.format_exc()}"
        )


if __name__ == '__main__':
    log_.info("文件alg_recsys_recall_24h_region.py:「24小时地域」 开始执行")
    h_timer_check()