Google Bigquery 中的嵌套字段使用 OFFSET 而不是 UNNEST

Using OFFSET instead of UNNEST for nested fields in Google Bigquery

向 GBQ 大师提问。

这是两个目的相同的查询

first

SELECT
  fullVisitorId AS userid,
  CONCAT(fullVisitorId, visitStartTime) AS session,
  visitStartTime + (hits[
  OFFSET(0)].time / 1000) AS eventtime,
  date,
  trafficSource.campaign,
  trafficSource.source,
  trafficSource.medium,
  trafficSource.adContent,
  trafficSource.adwordsClickInfo.campaignId,
  geoNetwork.region,
  geoNetwork.city,
  trafficSource.keyword,
  totals.visits AS visits,
  device.deviceCategory AS deviceType,
  hits[OFFSET(0)].eventInfo.eventAction,
  hits[OFFSET(0)].TRANSACTION.transactionId,
  hits[OFFSET(0)].TRANSACTION.transactionRevenue,
  SUBSTR(channelGrouping,0,3) AS newchannelGrouping
FROM
  `some_site.ga_sessions_*`
WHERE
  ARRAY_LENGTH(hits) > 0
  AND _table_suffix BETWEEN '20200201'
  AND '20200201'
  AND fullVisitorId IN (
  SELECT
    DISTINCT(fullVisitorId)
  FROM
    `some_site.ga_sessions_*`,
    UNNEST(hits) AS hits
  WHERE
    _table_suffix BETWEEN '20200201'
    AND '20200201'
    AND (hits.TRANSACTION.transactionId != 'None')
)

second

SELECT
  fullVisitorId AS userid,
  CONCAT(fullVisitorId, visitStartTime) AS session,
  visitStartTime + (hits.time / 1000) AS eventtime,
  date,
  trafficSource.campaign,
  trafficSource.source,
  trafficSource.medium,
  trafficSource.adContent,
  trafficSource.adwordsClickInfo.campaignId,
  geoNetwork.region,
  geoNetwork.city,
  trafficSource.keyword,
  totals.visits AS visits,
  device.deviceCategory AS deviceType,
  hits.eventInfo.eventAction,
  hits.TRANSACTION.transactionId,
  hits.TRANSACTION.transactionRevenue,
  SUBSTR(channelGrouping,0,3) AS newchannelGrouping
FROM
  `some_site.ga_sessions_*`, UNNEST(hits) hits
WHERE
  _table_suffix BETWEEN '20200201' AND '20200201'
  AND fullVisitorId IN (
    SELECT
      DISTINCT(fullVisitorId)
    FROM
      `some_site.ga_sessions_*`,
      UNNEST(hits) AS hits
    WHERE
      _table_suffix BETWEEN '20200201'
      AND '20200201'
      AND (hits.TRANSACTION.transactionId != 'None')
  )

第一个使用 OFFSET 从嵌套字段中提取数据。根据执行详细信息报告,查询需要大约 1.5 MB 的改组。

第二个查询使用 UNNEST 来访问嵌套数据。打乱后的字节数约为 (!) 75 MB

两种情况下处理的数据量相同。

现在,问题是:

这是否意味着根据 this article 优化插槽之间的通信,我应该使用 OFFSET 而不是 UNNEST 来获取存储在嵌套字段中的数据?

谢谢!

让我们考虑以下使用 BigQuery public 数据集的示例。

UNNEST - returns 6 个结果:

WITH t AS (SELECT * FROM `bigquery-public-data.google_analytics_sample.ga_sessions_20170801` WHERE visitId = 1501571504 )
SELECT h FROM t, UNNEST(hits) h

OFFSET - returns 1 个结果:

WITH t AS (SELECT * FROM `bigquery-public-data.google_analytics_sample.ga_sessions_20170801` WHERE visitId = 1501571504 )
SELECT hits[OFFSET(0)] FROM t

两个查询都引用了 GA public table 中的同一条记录。他们表明,使用带有 UNNEST 的连接将为数组中的每个元素带来一行,而使用 OFFSET(0) 只会为数组的第一个元素带来一行。

造成high data shuffling差异的原因是UNNEST执行了JOIN操作,需要以特定的方式组织数据。 OFFSET 方法仅采用数组的第一个元素。