[SPARK-30218] Columns used in inequality conditions for joins not resolved correctly in case of common lineage

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 2.3.4, 2.4.4
• Component/s: PySpark
• Labels: correctness
• Assignee: Unassigned
• Reporter: Francesco Cavrini
• Created: 11/Dec/19 09:56
• Updated: 24/Jan/20 17:10
• Resolved: 22/Jan/20 18:16

Description

When columns from different data-frames that have a common lineage are used in inequality conditions in joins, they are not resolved correctly. In particular, both the column from the left DF and the one from the right DF are resolved to the same column, thus making the inequality condition either always satisfied or always not-satisfied.

Minimal example to reproduce follows.

import pyspark.sql.functions as F

data = spark.createDataFrame([["id1", "A", 0], ["id1", "A", 1], ["id2", "A", 2], ["id2", "A", 3], ["id1", "B", 1] , ["id1", "B", 5], ["id2", "B", 10]], ["id", "kind", "timestamp"])

df_left = data.where(F.col("kind") == "A").alias("left")
df_right = data.where(F.col("kind") == "B").alias("right")

conds = [df_left["id"] == df_right["id"]]
conds.append(df_right["timestamp"].between(df_left["timestamp"], df_left["timestamp"] + 2))

res = df_left.join(df_right, conds, how="left")

The result is:

id	kind	timestamp	id	kind	timestamp
id1	A	0	id1	B	1
id1	A	0	id1	B	5
id1	A	1	id1	B	1
id1	A	1	id1	B	5
id2	A	2	id2	B	10
id2	A	3	id2	B	10

which violates the condition that the timestamp from the right DF should be between df_left["timestamp"] and df_left["timestamp"] + 2.

The plan shows the problem in the column resolution.

== Parsed Logical Plan ==
Join LeftOuter, ((id#0 = id#36) && ((timestamp#2L >= timestamp#2L) && (timestamp#2L <= (timestamp#2L + cast(2 as bigint)))))
:- SubqueryAlias `left`
:  +- Filter (kind#1 = A)
:     +- LogicalRDD [id#0, kind#1, timestamp#2L], false
+- SubqueryAlias `right`
   +- Filter (kind#37 = B)
      +- LogicalRDD [id#36, kind#37, timestamp#38L], false

Note, the columns used in the equality condition of the join have been correctly resolved.

Issue Links

• duplicates: SPARK-28344 fail the query if detect ambiguous self join
• relates to: SPARK-24780 DataFrame.column_name should resolve to a distinct ref