This may be surprising to regular pandas users and may lead to unexpected or silent errors.
Let’s construct an example dataframe to demonstrate the problem. The following dataframe
has two columns – column x has type integer, column y has type string. Column x
has an element 1 (integer) and column y has an element "1" (string).
df = spark.createDataFrame([(1, 'a'), (2, 'b'), (3, '1')], 'x INT, y STRING')
df.dtypes
# [('x', 'int'), ('y', 'string')]
df.show()
# +---+---+
# | x| y|
# +---+---+
# | 1| a|
# | 2| b|
# | 3| 1|
# +---+---+
Let’s say we want to count how many 1 (integer) values are in each column. We should get back
these results – column x should have one instance of 1 (integer) and column y should have
zero instances of 1 (integer). But only one of these happen.
df[df['x'].isin(1)].count() # 1 (correct)
df[df['y'].isin(1)].count() # 1 (incorrect)
We may suspect that the
pyspark.sql.Column.isin method
method has a bug. Let’s check another way without using .isin method.
df[df['x'] == 1].count() # 1 (correct)
df[df['y'] == 1].count() # 1 (incorrect)
We still get the same partially incorrect results. Let’s check something else – filter
for "1" (string) instead.
df[df['x'].isin("1")].count() # 1 (incorrect)
df[df['y'].isin("1")].count() # 1 (correct)
df[df['x'] == "1"].count() # 1 (incorrect)
df[df['y'] == "1"].count() # 1 (correct)
It appears that (1) .isin is not the reason and (2) spark does not respect data types when
filtering. Python is a dynamically typed language (though it now supports optional type checking).
Perhaps that is causing this. Let’s just use SQL (via python) instead and verify.
df.createOrReplaceTempView('dummy')
spark.sql('SELECT * FROM dummy WHERE x = 1').count() # 1 (correct)
spark.sql('SELECT * FROM dummy WHERE y = 1').count() # 1 (incorrect)
spark.sql('SELECT * FROM dummy WHERE x = "1"').count() # 1 (incorrect)
spark.sql('SELECT * FROM dummy WHERE y = "1"').count() # 1 (correct)
We get the same results back – data types are not respected.
Let’s investigate if this happens in a statically typed language like Scala.
import spark.implicits._
val df = Seq(
(1, "a"),
(2, "b"),
(3, "1")
).toDF("x", "y")
df.dtypes
// Array[(String, String)] = Array((x,IntegerType), (y,StringType))
df.filter($"x" === 1).count() // 1 (correct)
df.filter($"y" === 1).count() // 1 (incorrect)
df.filter($"x" === "1").count() // 1 (incorrect)
df.filter($"y" === 1).count() // 1 (correct)
Scala has the same problem!
Turns out we see the same behavior in at least some SQL systems. Let’s use SQL (using sqlite3)
without using spark at all.
import sqlite3
conn = sqlite3.connect(':memory:')
c = conn.cursor()
c.execute('CREATE TABLE dummy (x integer, y string)')
c.execute('INSERT INTO dummy VALUES (1, "a")')
c.execute('INSERT INTO dummy VALUES (2, "b")')
c.execute('INSERT INTO dummy VALUES (3, "1")')
conn.commit()
list(c.execute('SELECT * FROM dummy'))
list(c.execute('SELECT * FROM dummy WHERE x=1')) # [(1, 'a')] (correct)
list(c.execute('SELECT * FROM dummy WHERE y=1')) # [(3, 1)] (incorrect)
list(c.execute('SELECT * FROM dummy WHERE x="1"')) # [(1, 'a')] (incorrect)
list(c.execute('SELECT * FROM dummy WHERE y="1"')) # [(3, 1)] (correct)
conn.close()
Turns out SQL does not respect data types either!
pandas has this bug.Turns out that even pandas has this problem. For pyspark and SQL,
this problem appears to be a consistent design issue but for pandas this problem appears to be
bug instead of a design choice. I used pandas version 0.25.3 for this experiment.
pdf['x'].isin([1]).sum() # 1 (correct)
pdf['y'].isin([1]).sum() # 0 (correct)
pdf['x'].isin(["1"]).sum() # 1 (incorrect)
pdf['y'].isin(["1"]).sum() # 1 (correct)
Note how only the third line in the above snippet returns incorrect results. This indicates
that pandas ignores the data type sometimes but not always. Further, in the case of pandas
the problem is only in the .isin method and not in general (like in SQL and pyspark), as
shown by the following example that does not use .isin.
# Results are correct when we don't use pd.Series.isin() method
pdf['x'].apply(lambda x: x == 1).sum() # 1 (correct)
pdf['y'].apply(lambda y: y == 1).sum() # 0 (correct)
pdf['x'].apply(lambda x: x == "1").sum() # 0 (correct)
pdf['y'].apply(lambda y: y == "1").sum() # 1 (correct)
(pdf['x'] == 1).sum() # 1 (correct)
(pdf['y'] == 1).sum() # 0 (correct)
(pdf['x'] == "1").sum() # 0 (correct)
(pdf['y'] == "1").sum() # 1 (correct)
Looking even deeper into the pandas.Series.isin method, we see that it relies upon the
pandas.core.algorithms.isin function.
Lines 452-453
(in pandas version 0.25.3) contain the following code (comments are mine):
# Lines 452-453 in pandas 0.25.3
comps, dtype, _ = _ensure_data(comps) # comps = elements of the column
values, _, _ = _ensure_data(values, dtype=dtype) # values = list of values passed to `.isin`
As you can see, values get type casted into the dtype of the column. We can verify this
by actually running the code ourselves.
from pandas.core.algorithms import _ensure_data
comps = pdf['x']
values = ["1"] # str
comps, dtype, _ = _ensure_data(comps)
print(dtype) # int64
values, _, _ = _ensure_data(values, dtype=dtype)
print(values) # [1]
print(values.dtype) # int64
The reason why this does not affect column y in pdf above is because _ensure_data returns
dtype=object for column y. Searching pandas GitHub issues, we see that
this bug
has been brought up before but was somehow ignored.
Spark and SQL ignore data types when filtering on equality. This seems to be a design issue and is consistent throughout (or at least as far as I can see). Pandas, on the other hand, exhibits this problem inconsistently which may lead to complacency.
For end users, the best way to prevent this mistake is to always manually ensure that the data types of the column match the data type of every value in a list values used for equality filtering.