How to join on multiple columns in Pyspark?
Question:
I am using Spark 1.3 and would like to join on multiple columns using python interface (SparkSQL)
The following works:
I first register them as temp tables.
numeric.registerTempTable("numeric")
Ref.registerTempTable("Ref")
test = numeric.join(Ref, numeric.ID == Ref.ID, joinType='inner')
I would now like to join them based on multiple columns.
I get SyntaxError
: invalid syntax with this:
test = numeric.join(Ref,
numeric.ID == Ref.ID AND numeric.TYPE == Ref.TYPE AND
numeric.STATUS == Ref.STATUS , joinType='inner')
Answers:
You should use &
/ |
operators and be careful about operator precedence (==
has lower precedence than bitwise AND
and OR
):
df1 = sqlContext.createDataFrame(
[(1, "a", 2.0), (2, "b", 3.0), (3, "c", 3.0)],
("x1", "x2", "x3"))
df2 = sqlContext.createDataFrame(
[(1, "f", -1.0), (2, "b", 0.0)], ("x1", "x2", "x3"))
df = df1.join(df2, (df1.x1 == df2.x1) & (df1.x2 == df2.x2))
df.show()
## +---+---+---+---+---+---+
## | x1| x2| x3| x1| x2| x3|
## +---+---+---+---+---+---+
## | 2| b|3.0| 2| b|0.0|
## +---+---+---+---+---+---+
An alternative approach would be:
df1 = sqlContext.createDataFrame(
[(1, "a", 2.0), (2, "b", 3.0), (3, "c", 3.0)],
("x1", "x2", "x3"))
df2 = sqlContext.createDataFrame(
[(1, "f", -1.0), (2, "b", 0.0)], ("x1", "x2", "x4"))
df = df1.join(df2, ['x1','x2'])
df.show()
which outputs:
+---+---+---+---+
| x1| x2| x3| x4|
+---+---+---+---+
| 2| b|3.0|0.0|
+---+---+---+---+
With the main advantage being that the columns on which the tables are joined are not duplicated in the output, reducing the risk of encountering errors such as org.apache.spark.sql.AnalysisException: Reference 'x1' is ambiguous, could be: x1#50L, x1#57L.
Whenever the columns in the two tables have different names, (let’s say in the example above, df2
has the columns y1
, y2
and y4
), you could use the following syntax:
df = df1.join(df2.withColumnRenamed('y1','x1').withColumnRenamed('y2','x2'), ['x1','x2'])
test = numeric.join(Ref,
on=[
numeric.ID == Ref.ID,
numeric.TYPE == Ref.TYPE,
numeric.STATUS == Ref.STATUS
], how='inner')
You can also provide a list of strings, if the column names are the same.
df1 = sqlContext.createDataFrame(
[(1, "a", 2.0), (2, "b", 3.0), (3, "c", 3.0)],
("x1", "x2", "x3"))
df2 = sqlContext.createDataFrame(
[(1, "f", -1.0), (2, "b", 0.0)], ("x1", "x2", "x3"))
df = df1.join(df2, ["x1","x2"])
df.show()
+---+---+---+---+
| x1| x2| x3| x3|
+---+---+---+---+
| 2| b|3.0|0.0|
+---+---+---+---+
Another way to go about this, if column names are different and if you want to rely on column name strings is the following:
df1 = sqlContext.createDataFrame(
[(1, "a", 2.0), (2, "b", 3.0), (3, "c", 3.0)],
("x1", "x2", "x3"))
df2 = sqlContext.createDataFrame(
[(1, "f", -1.0), (2, "b", 0.0)], ("y1", "y2", "y3"))
df = df1.join(df2, (col("x1")==col("y1")) & (col("x2")==col("y2")))
df.show()
+---+---+---+---+---+---+
| x1| x2| x3| y1| y2| y3|
+---+---+---+---+---+---+
| 2| b|3.0| 2| b|0.0|
+---+---+---+---+---+---+
This is useful if you want to reference column names dynamically and also in instances where there is a space in the column name and you cannot use the df.col_name
syntax. You should look at changing the column name in that case anyway though.
I am using Spark 1.3 and would like to join on multiple columns using python interface (SparkSQL)
The following works:
I first register them as temp tables.
numeric.registerTempTable("numeric")
Ref.registerTempTable("Ref")
test = numeric.join(Ref, numeric.ID == Ref.ID, joinType='inner')
I would now like to join them based on multiple columns.
I get SyntaxError
: invalid syntax with this:
test = numeric.join(Ref,
numeric.ID == Ref.ID AND numeric.TYPE == Ref.TYPE AND
numeric.STATUS == Ref.STATUS , joinType='inner')
You should use &
/ |
operators and be careful about operator precedence (==
has lower precedence than bitwise AND
and OR
):
df1 = sqlContext.createDataFrame(
[(1, "a", 2.0), (2, "b", 3.0), (3, "c", 3.0)],
("x1", "x2", "x3"))
df2 = sqlContext.createDataFrame(
[(1, "f", -1.0), (2, "b", 0.0)], ("x1", "x2", "x3"))
df = df1.join(df2, (df1.x1 == df2.x1) & (df1.x2 == df2.x2))
df.show()
## +---+---+---+---+---+---+
## | x1| x2| x3| x1| x2| x3|
## +---+---+---+---+---+---+
## | 2| b|3.0| 2| b|0.0|
## +---+---+---+---+---+---+
An alternative approach would be:
df1 = sqlContext.createDataFrame(
[(1, "a", 2.0), (2, "b", 3.0), (3, "c", 3.0)],
("x1", "x2", "x3"))
df2 = sqlContext.createDataFrame(
[(1, "f", -1.0), (2, "b", 0.0)], ("x1", "x2", "x4"))
df = df1.join(df2, ['x1','x2'])
df.show()
which outputs:
+---+---+---+---+
| x1| x2| x3| x4|
+---+---+---+---+
| 2| b|3.0|0.0|
+---+---+---+---+
With the main advantage being that the columns on which the tables are joined are not duplicated in the output, reducing the risk of encountering errors such as org.apache.spark.sql.AnalysisException: Reference 'x1' is ambiguous, could be: x1#50L, x1#57L.
Whenever the columns in the two tables have different names, (let’s say in the example above, df2
has the columns y1
, y2
and y4
), you could use the following syntax:
df = df1.join(df2.withColumnRenamed('y1','x1').withColumnRenamed('y2','x2'), ['x1','x2'])
test = numeric.join(Ref,
on=[
numeric.ID == Ref.ID,
numeric.TYPE == Ref.TYPE,
numeric.STATUS == Ref.STATUS
], how='inner')
You can also provide a list of strings, if the column names are the same.
df1 = sqlContext.createDataFrame(
[(1, "a", 2.0), (2, "b", 3.0), (3, "c", 3.0)],
("x1", "x2", "x3"))
df2 = sqlContext.createDataFrame(
[(1, "f", -1.0), (2, "b", 0.0)], ("x1", "x2", "x3"))
df = df1.join(df2, ["x1","x2"])
df.show()
+---+---+---+---+
| x1| x2| x3| x3|
+---+---+---+---+
| 2| b|3.0|0.0|
+---+---+---+---+
Another way to go about this, if column names are different and if you want to rely on column name strings is the following:
df1 = sqlContext.createDataFrame(
[(1, "a", 2.0), (2, "b", 3.0), (3, "c", 3.0)],
("x1", "x2", "x3"))
df2 = sqlContext.createDataFrame(
[(1, "f", -1.0), (2, "b", 0.0)], ("y1", "y2", "y3"))
df = df1.join(df2, (col("x1")==col("y1")) & (col("x2")==col("y2")))
df.show()
+---+---+---+---+---+---+
| x1| x2| x3| y1| y2| y3|
+---+---+---+---+---+---+
| 2| b|3.0| 2| b|0.0|
+---+---+---+---+---+---+
This is useful if you want to reference column names dynamically and also in instances where there is a space in the column name and you cannot use the df.col_name
syntax. You should look at changing the column name in that case anyway though.