[SPARK-6072] Enable hash joins for null-safe equality predicates

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 1.2.1
• Component/s: SQL
• Assignee: Unassigned
• Reporter: Dima Zhiyanov
• Created: 28/Feb/15 00:53
• Updated: 12/Dec/22 18:10
• Resolved: 20/Mar/17 03:43

Description

Currently joins such as
A join B on A.x = B.x AND A.y <=> B.y
are evaluated as hash join on just x followed by filter on y.
This causes a skew problem (very long join) when a particular value of x has a high cardinality even though (x, y) is evenly distributed

Can we implement is as a hash join on (X, Option(Y))? This will eliminate the skew in this case

Imagine a join:
People as p1 join People as p2 on p1.name = p2.name and p1.address <=> p2.address

(very small percentage of people has unknown address)

This causes a skewed join on popular names such as "Mary Brown" if we hash on names alone, but will not cause a skew if we hash on (Name, Option(Address))

Issue Links

• duplicates: SPARK-11111 Fast null-safe join