google-bigquery - Bigqueryの移動平均

Question

bigqueryで移動平均を計算する方法を教えてもらえますか？

これが私がmysqlスタイルで必要なものです。

SELECT T1.id, T1.value_column1, avg(T2.value_column1)
FROM table1 T1
INNER JOIN table1 T2 ON T2.Id BETWEEN T1.Id-19 AND T1.Id

Answer 1

同じことを行うことができますが、BigQueryでは同等の結合しか許可されていないため、少し手間がかかります。これは、公的な出生率サンプルから出生時体重の6か月移動平均を計算する例です。

SELECT 
  --Convert months-since-year-0 back to year, month
  INTEGER(month / 12) as year, 
  month % 12 as month, 
  avg 
FROM (
  SELECT month, 
    -- Note that this average is the average over all of the data in the 
    -- last 6 months, not an average over the avg values for the last 6 months.
    -- It is easy to compute the latter, if that is what is desired -- just 
    -- compute the average in the inner select, and take the average of those
    -- here.
     SUM(total_weight_per_month) / SUM(records_per_month) as avg
  FROM (
    SELECT 
       -- Note we use t2.month here since that is what is compared against
       -- 6 different t1 months.
       t2.month as month, 
       t1.records_per_month as records_per_month,     
       t1.total_weight_per_month as total_weight_per_month 
    FROM (
      SELECT month,
        COUNT( weight_pounds ) as records_per_month, 
        SUM( weight_pounds)  as total_weight_per_month,
        -- This active field is the key that lets us join all of the
        -- values against the values in the date subselect.
        1 AS active
      FROM (
        SELECT 
           -- Convert year and month fields to a single value that
           -- has the number of months since year 0. This will allow
           -- us to do math on the dates.
          year * 12 + month AS month, 
          weight_pounds 
        FROM [publicdata:samples.natality] 
        WHERE weight_pounds > 0)
      GROUP BY month) as t1
    JOIN  
      -- We join the weights per month agsint a subselect that contains
      -- all months.
      (SELECT month, 1 as active 
      FROM
        (SELECT 
           year * 12 + month AS month, 
        FROM [publicdata:samples.natality])
      GROUP BY month) as t2
    ON t1.active = t2.active 
    -- Here is where we get the moving average -- we basically take the month 
    -- value from t1 and make it apply for 6 months.
    WHERE t1.month >= t2.month && t1.month - 6 < t2.month)
  GROUP BY month
  ORDER BY month desc)

Answer 2

更新されたより効率的な回答については、https：//stackoverflow.com/a/24943950/132438をご覧ください。

---

新しいLAG（）およびLEAD（）ウィンドウ関数を確認します。自己結合を必要とせずに、結果セットをトラバースできます。

https://developers.google.com/bigquery/docs/query-reference#windowfunctions

JOIN EACHの別のオプション（中間ステップで非常に大量のデータが生成される可能性があるため、これは遅くなりすぎる可能性があります）：

SELECT a.SensorId SensorId, a.Timestamp, AVG(b.Data) AS avg_prev_hour_load
FROM (
  SELECT * FROM [io_sensor_data.moscone_io13]
  WHERE SensorId = 'XBee_40670EB0/mic') a
JOIN EACH [io_sensor_data.moscone_io13] b
ON a.SensorId = b.SensorId
WHERE b.Timestamp BETWEEN (a.Timestamp - 3600000) AND a.Timestamp
GROUP BY SensorId, a.Timestamp;

（Joe CelkoのSQLの問題に基づく）

Answer 3

次の「Times」テーブルを作成しました。

Table Details: Dim_Periods
Schema
Date    TIMESTAMP   
Year    INTEGER         
Month   INTEGER         
day         INTEGER         
QUARTER INTEGER     
DAYOFWEEK   INTEGER     
MonthStart  TIMESTAMP   
MonthEnd    TIMESTAMP   
WeekStart   TIMESTAMP   
WeekEnd TIMESTAMP   
Back30Days  TIMESTAMP   -- the date 30 days before "Date"
Back7Days   TIMESTAMP   -- the date 7 days before "Date"

そして、私はそのようなクエリを使用して「ランニングサム」を処理します

SELECT Date,Count(*) as MovingCNT
FROM

(SELECT Date,
                Back7Days 
                    FROM DWH.Dim_Periods  
                 where Date < timestamp(current_date()) AND
                             Date >= (DATE_ADD (CURRENT_TIMESTAMP(), -5, 'month'))
                )P
                CROSS JOIN EACH
    (SELECT repository_url,repository_created_at
    FROM publicdata:samples.github_timeline
                ) L
        WHERE timestamp(repository_created_at)>= Back7Days 
              AND timestamp(repository_created_at)<= Date

GROUP EACH BY Date

「Monthtodate」、「Week to Date」、「30 days back」などの集計にも使用できることに注意してください。ただし、パフォーマンスは最高ではなく、デカルト座標のため、大規模なデータセットではクエリに時間がかかる場合があります。参加してください。これがお役に立てば幸いです。