NTILE 與 CUME_DIST

Snowflake 的 Window Functions

Jake Roach

Field Data engineer

建立分桶群組

如何依會員的運動表現分群，對應推廣合適的課程？

             member_id  |  gym_location  |  calories_burned  |  marketing_group
            ----------- | -------------- | ----------------- | -----------------
               m_192    |    Miami       |         45        |         1
               m_74     |    Miami       |         59        |         1
               m_233    |    Portland    |         60        |         1

               m_14     |    Cleveland   |         72        |         2
               m_346    |    Portland    |         77        |         2
               m_289    |    Cleveland   |         81        |         2

               m_565    |    Miami       |        1085       |         50
                                        ...

NTILE

SELECT
    <fields>,
    <2>,
    <1>,


    NTILE(<n>) OVER(
        PARTITION BY <2>
        ORDER BY <1>
    )


...;

NTILE 用來建立 N 個大小相近的「分桶」

<n>：分桶數量

<1>：用來分桶的欄位

<2>：用 PARTITION BY 平均分配記錄的欄位

為健身資料分桶

SELECT
    member_id,
    gym_location,
    calories_burned,

    -- Create 50 equally-sized buckets of data

    NTILE(50) OVER(
        ORDER BY calories_burned   -- 決定每個分桶的記錄
    ) AS marketing_group


FROM FITNESS.workouts
ORDER BY marketing_group, calories_burned;  -- 排序最終結果集

為健身資料分桶

         member_id  |  gym_location  |  calories_burned  |  marketing_group
        ----------- | -------------- | ----------------- | -----------------
           m_192    |    Miami       |         45        |         1
           m_74     |    Miami       |         59        |         1
           m_233    |    Portland    |         60        |         1

           m_14     |    Cleveland   |         72        |         2
           m_346    |    Portland    |         77        |         2
           m_289    |    Cleveland   |         81        |         2

           m_565    |    Miami       |        1085       |         50
                                        ...

平均分配的健身分桶

SELECT
    member_id,
    gym_location,
    calories_burned,

    NTILE(50) OVER(

        -- 依各 gym_location 在分桶內平均分配記錄
        PARTITION BY gym_location

        ORDER BY calories_burned
    ) AS marketing_group

FROM FITNESS.workouts
ORDER BY marketing_group, calories_burned;

平均分配的健身分桶

         member_id  |  gym_location  |  calories_burned  |  marketing_group
        ----------- | -------------- | ----------------- | -----------------
           m_192    |    Miami       |         45        |         1
           m_233    |    Portland    |         60        |         1
           m_14     |    Cleveland   |         72        |         1

           m_74     |    Miami       |         59        |         2
           m_346    |    Portland    |         77        |         2
           m_289    |    Cleveland   |         81        |         2

                                    ...

理解分布

各會員的卡路里消耗分布是什麼樣貌？

特定的一次運動在此分布中的位置在哪？

有多少比例的會員消耗的卡路里小於或等於某會員？

  member_id  |  cals_burned  |   cd   
 ----------- | ------------- | ------
    m_192    |       45      |  .016
    m_74     |       59      |  .033
    m_233    |       60      |  .049
    m_14     |       72      |  .066
    m_346    |       77      |  .082
    m_289    |       81      |  .098

                    ....

    m_565    |      1085     |  1.000

CUME_DIST

SELECT
    member_id,
    gym_location,
    calories_burned,

    CUME_DIST() OVER(
        PARTITION BY gym_location,   -- 為各地點建立分布
        ORDER BY calories_burned
    ) AS cd

FROM FITNESS.workouts

ORDER BY gym_location, cd;  -- 排序最終結果集

CUME_DIST

SELECT
    <fields>,
    <1>,
    <2>,


    CUME_DIST() OVER(
        PARTITION BY <1>
        ORDER BY <2>
    )


...;

將每筆記錄與該欄位的分布比較，得到「累積分布」

<1>：決定評估視窗的欄位

<2>：用來建立分布的欄位

有多少比例的記錄小於或等於此筆？

CUME_DIST

               member_id  |  gym_location  | calories_burned  |   cd   
              ----------- | -------------- | ---------------- | -------
                 m_192    |    Miami       |        45        |  .033
                 m_74     |    Miami       |        59        |  .066
                 m_288    |    Miami       |        83        |  .098
                 m_541    |    Miami       |        85        |  .131

                                          ...

                 m_233    |    Portland    |        60        |  .071
                 m_346    |    Portland    |        77        |  .142

                                          ...

一起來練習吧！

Snowflake 的 Window Functions