How Do I Do When Data Skew Occurs During the Execution of a SQL Job?

What Is Data Skew?

Figure 1 Data skew example

Common Data Skew Scenarios

• Group By aggregation skew

  During the execution of Group By aggregation, if some grouping keys have significantly more data than others, the larger groups will consume more compute resources and time during the aggregation process, resulting in slower processing speeds and data skew.

• JOIN operation skew

  During table JOIN operations, if the keys involved in the JOIN are unevenly distributed in one of the tables, a large amount of data will be concentrated in a few tasks while others have already completed, causing data skew.

Solution for Group By Data Skew

SELECT
  a.Key,
  SUM(a.sum) AS Cnt
FROM
  (
    SELECT
      Key,
      count(*) AS sum
    FROM
      tbl
    GROUP BY
      Key,
      CASE
        WHEN KEY = 'Key01' THEN floor(random () * 200)
        ELSE 0
      END
  ) a
GROUP BY
  a.Key;

Solution for JOIN Data Skew

1. Log in to the DLI management console. Choose Job Management > SQL Jobs in the navigation pane. On the displayed page, locate the job you want to modify and click Edit in the Operation column to switch to the SQL Editor page.
2. On the SQL editor page, click Set Property and add the following Spark parameters through the Settings pane:

   The string followed by the colons (:) are the configuration parameters, and the strings following the colons are the values.

   spark.sql.enableToString:false
   spark.sql.adaptive.join.enabled:true
   spark.sql.adaptive.enabled:true
   spark.sql.adaptive.skewedJoin.enabled:true
   spark.sql.adaptive.enableToString:false
   spark.sql.adaptive.skewedPartitionMaxSplits:10
   
   

   Note

   spark.sql.adaptive.skewedPartitionMaxSplits indicates the maximum number of tasks for processing a skewed partition. The default value is 5, and the maximum value is 10. This parameter is optional.

3. Click Execute to run the job again.