[BUG] String match returns incorrect result in some cases

[ayushdg]

Describe the bug
series.str.match returns incorrect results in the case of some regular expressions.

Steps/Code to reproduce bug

import cudf

regex = "^\\^|()\\-*\\[\\]\\$$"
pdf = pd.DataFrame({"a": ["a normal string", "%_%", "^|()-*[]$"]})
df = cudf.from_pandas(pdf)

print(df.a.str.match(regex))

0    True
1    True
2    True
Name: a, dtype: bool

Expected behavior
In the above example

print(pdf.a.str.match(regex))

0    False
1    False
2     True
Name: a, dtype: bool

Environment overview (please complete the following information)

• Environment location: Bare-metal
• Method of cuDF install: conda
  • If method of install is [Docker], provide docker pull & docker run commands used

Environment details
Please run and paste the output of the cudf/print_env.sh script here, to gather any other relevant environment details

Additional context
Came up while adding gpu tests for string functionality in dask-sql: dask-contrib/dask-sql#256

[davidwendt]

It looks like there are a lot of unescaped regex characters in the pattern that may be confusing the code. Here is the documentation that includes the meaning for the special regex characters. Looks like if you escape them all, you will get the correct result.

import cudf
regex = "^\\^\\|\\(\\)\\-\\*\\[\\]\\$$"
s = cudf.Series(["a normal string", "%_%", "^|()-*[]$"])
print(s.str.match(regex))

0    False
1    False
2     True
dtype: bool

[ayushdg]

Thanks for the context and link.
Are the regex rules for cudf string pattern matching slightly different than the standard python regex rules? (In terms of requiring to escape certain chars)

Trying to evaluate if dask-sql will need a special case to handle regexes differently for gpu dataframes vs pandas df's.

[davidwendt]

I would expect that all regex characters would need to be escaped if they are to be matched literally in a target string. Perhaps python has done some work to guess when the escape was intended? I don't have a clue how they would know to do that.

[andygrove]

I am also running into this issue. Here is one example:

Python

>>> pattern = re.compile('a[-+]')
>>> values = ['a-', 'a+', 'a']
>>> [pattern.search(x) for x in values]
[<re.Match object; span=(0, 2), match='a-'>, <re.Match object; span=(0, 2), match='a+'>, None]
>>> 

cuDF with escaping (correct results)

>>> cudf.Series(['a+', 'a-', 'a']).str.contains('a[\-\+]', regex=True)
0     True
1     True
2    False
dtype: bool

cuDF without escaping

>>> cudf.Series(['a+', 'a-', 'a']).str.contains('a[-+]', regex=True)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/andy/miniconda3/envs/rapids-21.10/lib/python3.7/site-packages/cudf/core/column/string.py", line 743, in contains
    result_col = libstrings.contains_re(self._column, pat)
  File "cudf/_lib/strings/contains.pyx", line 29, in cudf._lib.strings.contains.contains_re
RuntimeError: cuDF failure at: ../src/strings/regex/regcomp.cpp:397: invalid regex pattern: nothing to repeat at position 3

[davidwendt]

The - inside [] is expecting a character range like [a-z] or [1-9] but is missing the first range character before the -.
So the - will need to be escaped here. Looks like + is a special character too.

[davidwendt]

Documenting this as undefined behavior here https://github.com/rapidsai/cudf/pull/9473/files#diff-5c59d57ecc67d4a26c406b9f0a4976a9c49c74c7c9ac8e9e4d3ba947fa065cc6R17-R23