会话持续时间 - 谷歌分析和 Firebase 报告之间的巨大差异

Posted 2023-03-24

【中文标题】会话持续时间 - 谷歌分析和 Firebase 报告之间的巨大差异

【英文标题】：Session Duration - large discrepancy between google analytics and firebase reporting

【发布时间】：2019-11-02 00:59:41

【问题描述】：

我们目前正在从谷歌分析（移动数据）迁移到 Firebase。

firebase 中的数据正在流式传输到 BigData，我们使用以下查询来计算会话持续时间。

当我针对这些会话生成报告时，通过 Firebase 报告的平均会话持续时间大约是 Google 分析报告的平均会话持续时间的 4 倍。

我在各种设备（android、ios）上都有同样的问题，并且同时向谷歌分析和 firebase 发送相同的自定义事件

使用 session_start 事件或 unnest user_properties 并直接使用 ga_session_number 来代替给出相同的结果。

with timeline as
(
  select 
    geo.city as geo_city
    ,device.operating_system as device_category
    , user_pseudo_id
    , event_name
    , event_timestamp
    , lag(event_timestamp, 1) over (partition by user_pseudo_id order by event_timestamp) as prev_event_timestamp
  from 
    `projid.analytics_xxx.events_*`
) 
, session_timeline as 
(
  select
    event_name
    , geo_city
    , device_category
    , user_pseudo_id
    , event_timestamp
    , prev_event_timestamp
    , case 
        when 
           event_timestamp - prev_event_timestamp >= (30*60*1000*1000)
             or
           prev_event_timestamp is null 
          then 1
          else 0 
      end as is_new_session_flag 
  from 
    timeline
)
, marked_sessions as
(
  select 
    device_category
    , geo_city
    , user_pseudo_id
    , event_timestamp
    , prev_event_timestamp
    , sum(is_new_session_flag) over (partition by user_pseudo_id order by event_timestamp) AS user_session_id
    , min(prev_event_timestamp) over (partition by user_pseudo_id order by event_timestamp) AS first_prev_event_timestamp

  from session_timeline
)
, measured_sessions as
(
  select
    device_category
    , geo_city
    , user_pseudo_id
    , user_session_id
    , FORMAT_TIMESTAMP("%F", TIMESTAMP_MICROS(first_prev_event_timestamp)) as first_time_date
    , FORMAT_TIMESTAMP("%F", TIMESTAMP_MICROS(min(event_timestamp))) as date
    , round((max(event_timestamp) - min(event_timestamp))/ (1000 * 1000), 2) as session_duration_seconds
    , round((max(event_timestamp) - min(event_timestamp))/ (1000 * 1000 * 60), 2) as session_duration_minutes
    , round((max(event_timestamp) - min(event_timestamp))/ (1000 * 1000 * 60 * 60), 2) as session_duration_hours 
  from 
    marked_sessions
  group by
    user_pseudo_id
    , user_session_id
    , device_category
    , geo_city
    , first_prev_event_timestamp
  having 
    -- let's count only sessions longer than 10 seconds
    session_duration_seconds >= 10
)
select 
  *
from 
  measured_sessions
 order by date desc, user_pseudo_id desc, user_session_id desc

不仅平均会话时长存在显着差异，甚至会话数也相差甚远。

【参考方案1】：

您的结果是否经过抽样？通常reporting discrepancies in Google Analytics 是由于采样。尝试导入按较少维度分组的数据或使用较短的时间段。