Question

有没有更好的方法来使用Apache Hive来对我的日志数据进行会话？我不确定我是否会以最佳方式执行此操作：

日志数据存储在序列文件中;单个日志条目是JSON字符串;例如：

{"source": {"api_key": "app_key_1", "user_id": "user0"}, "events": [{"timestamp": 1330988326, "event_type": "high_score", "event_params": {"score": "1123", "level": "9"}}, {"timestamp": 1330987183, "event_type": "some_event_0", "event_params": {"some_param_00": "val", "some_param_01": 100}}, {"timestamp": 1330987775, "event_type": "some_event_1", "event_params": {"some_param_11": 100, "some_param_10": "val"}}]}

格式化，如下所示：

{'source': {'api_key': 'app_key_1', 'user_id': 'user0'},
 'events': [{'event_params': {'level': '9', 'score': '1123'},
             'event_type': 'high_score',
             'timestamp': 1330988326},
            {'event_params': {'some_param_00': 'val', 'some_param_01': 100},
             'event_type': 'some_event_0',
             'timestamp': 1330987183},
            {'event_params': {'some_param_10': 'val', 'some_param_11': 100},
             'event_type': 'some_event_1',
             'timestamp': 1330987775}]
}

'source'包含一些关于'events'中包含的事件来源的信息（user_id和api_key）; 'events'包含源生成的事件列表;每个事件都有'event_params'，'event_type'和'timestamp'（时间戳是GMT中的Unix时间戳）。请注意，单个日志条目和日志条目中的时间戳可能不正常。

请注意，我受限制，以至于我无法更改日志格式，最初无法将数据记录到已分区的单独文件中（尽管我可以在记录数据后使用Hive执行此操作）等。

最后，我想要一个会话表，其中一个会话与app（api_k）和用户相关联，并且具有开始时间和会话长度（或结束时间）;会话被分开，对于给定的应用和用户，事件之间会出现30分钟或更长时间的差距。

我的解决方案执行以下操作（Hive脚本和python转换脚本在下面;看起来似乎不会显示SerDe源，但请告诉我它是否有用）：

[1]以非规范化格式

将数据加载到log_entry_tmp中

[2]将数据分解为log_entry，以便例如上面的单个条目现在有多个条目：

{"source_api_key":"app_key_1","source_user_id":"user0","event_type":"high_score","event_params":{"score":"1123","level":"9"},"event_timestamp":1330988326}
{"source_api_key":"app_key_1","source_user_id":"user0","event_type":"some_event_0","event_params":{"some_param_00":"val","some_param_01":"100"},"event_timestamp":1330987183}
{"source_api_key":"app_key_1","source_user_id":"user0","event_type":"some_event_1","event_params":{"some_param_11":"100","some_param_10":"val"},"event_timestamp":1330987775}

[3]转换并将数据写入session_info_0，其中每个条目包含事件'app_id，user_id和timestamp

[4]转换并将数据写入session_info_1，其中条目按app_id，user_id，event_timestamp排序;每个条目都包含一个session_id; python tranform脚本查找拆分，并将数据分组到会话

[5]转换并将最终会话数据写入session_info_2;会话的应用+用户，开始时间和长度（秒）

[Hive script]

drop table if exists app_info;
create external table app_info ( app_id int, app_name string, api_k string )
location '${WORK}/hive_tables/app_info';

add jar ../build/our-serdes.jar;

-- [1] load the data into log_entry_tmp, in a denormalized format

drop table if exists log_entry_tmp;
create external table log_entry_tmp
row format serde 'com.company.TestLogSerde'
location '${WORK}/hive_tables/test_logs';

drop table if exists log_entry;
create table log_entry (
    entry struct<source_api_key:string,
                 source_user_id:string,
                 event_type:string,
                 event_params:map<string,string>,
                 event_timestamp:bigint>);

-- [2] explode the data into log_entry

insert overwrite table log_entry
select explode (trans0_list) t
from log_entry_tmp;

drop table if exists session_info_0;
create table session_info_0 (
    app_id string,
    user_id string,
    event_timestamp bigint
);

-- [3] transform and write data into session_info_0, where each entry contains events' app_id, user_id, and timestamp

insert overwrite table session_info_0
select ai.app_id, le.entry.source_user_id, le.entry.event_timestamp
from log_entry le
join app_info ai on (le.entry.source_api_key = ai.api_k);

add file ./TestLogTrans.py;

drop table if exists session_info_1;
create table session_info_1 (
    session_id string,
    app_id string,
    user_id string,
    event_timestamp bigint,
    session_start_datetime string,
    session_start_timestamp bigint,
    gap_secs int
);

-- [4] tranform and write data into session_info_1, where entries are ordered by app_id, user_id, event_timestamp ; and each entry contains a session_id ; the python tranform script finds the splits, and groups the data into sessions

insert overwrite table session_info_1
select
    transform (t.app_id, t.user_id, t.event_timestamp)
        using './TestLogTrans.py'
        as (session_id, app_id, user_id, event_timestamp, session_start_datetime, session_start_timestamp, gap_secs)
from
    (select app_id as app_id, user_id as user_id, event_timestamp as event_timestamp from session_info_0 order by app_id, user_id, event_timestamp ) t;

drop table if exists session_info_2;
create table session_info_2 (
    session_id string,
    app_id string,
    user_id string,
    session_start_datetime string,
    session_start_timestamp bigint,
    len_secs int
);

-- [5] transform and write final session data to session_info_2 ; the sessions' app + user, start time, and length in seconds

insert overwrite table session_info_2
select session_id, app_id, user_id, session_start_datetime, session_start_timestamp, sum(gap_secs)
from session_info_1
group by session_id, app_id, user_id, session_start_datetime, session_start_timestamp;

[TestLogTrans.py]

#!/usr/bin/python

import sys, time

def buildDateTime(ts):

    return time.strftime('%Y-%m-%d %H:%M:%S', time.gmtime(ts))

curGroup = None
prevGroup = None
curSessionStartTimestamp = None
curSessionStartDatetime = None
prevTimestamp = None

for line in sys.stdin.readlines():

    fields = line.split('\t')
    if len(fields) != 3:
        raise Exception('fields = %s', fields)

    app_id = fields[0]
    user_id = fields[1]
    event_timestamp = int(fields[2].strip())

    curGroup = '%s-%s' % (app_id, user_id)
    curTimestamp = event_timestamp

    if prevGroup == None:
        prevGroup = curGroup
        curSessionStartTimestamp = curTimestamp
        curSessionStartDatetime = buildDateTime(curSessionStartTimestamp)
        prevTimestamp = curTimestamp

    isNewGroup = (curGroup != prevGroup)

    gapSecs = 0 if isNewGroup else (curTimestamp - prevTimestamp)

    isSessionSplit = (gapSecs >= 1800)

    if isNewGroup or isSessionSplit:
        curSessionStartTimestamp = curTimestamp
        curSessionStartDatetime = buildDateTime(curSessionStartTimestamp)

    session_id = '%s-%s-%d' % (app_id, user_id, curSessionStartTimestamp)

    print '%s\t%s\t%s\t%d\t%s\t%d\t%d' % (session_id, app_id, user_id, curTimestamp, curSessionStartDatetime, curSessionStartTimestamp, gapSecs)

    prevGroup = curGroup
    prevTimestamp = curTimestamp

Answer 1

我认为你可以轻松地删除第3步，并将你在那里使用的查询作为子查询添加到第4步中的from子句。物理化转换，似乎没有给你任何东西。

否则我认为你想在这里实现的目标，这似乎是一种合理的方法。

可能的步骤2您可以使用自定义映射器，将输出作为自定义缩减器传递到步骤4（步骤3作为子查询内置）。这会将mapreduce作业减少1，因此可以节省大量时间。

使用Apache Hive会话化我的日志数据的更好方法是什么？

1 个答案: