spark: How does salting work in dealing with skewed data

问题

I have a skewed data in a table which is then compared with other table that is small. I understood that salting works in case of joins- that is a random number is appended to keys in big table with skew data from a range of random data and the rows in small table with no skew data are duplicated with the same range of random numbers. Hence the the matching happens because there will be a hit in one among the duplicate values for particular slated key of skewed able I also read that salting is helpful while performing groupby. My question is when random numbers are appended to the key doesn't it break the group? If if does then the meaning of group by operation has changed.

回答1:

My question is when random numbers are appended to the key doesn't it break the group?

Well, it does, to mitigate this you could run group by operation twice. Firstly with salted key, then remove salting and group again. The second grouping will take partially aggregated data, thus significantly reduce skew impact.

E.g.

import org.apache.spark.sql.functions._

df.withColumn("salt", (rand * n).cast(IntegerType))
  .groupBy("salt", groupByFields)
  .agg(aggFields)
  .groupBy(groupByFields)
  .agg(aggFields)

来源：https://stackoverflow.com/questions/58110207/spark-how-does-salting-work-in-dealing-with-skewed-data