[BUG] Possible race condition in regular expression support for octal digits

[andygrove]

Describe the bug

I am seeing odd results with regular expressions containing octal digits. During testing I seem to be able to trigger a race condition where the first call to containsRe will produce false and subsequent calls will return true. Here is a test that reproduces the issue (to be added to RegularExpressionTranspilerSuite).

test("possible race condition") {
  for (_ <- 0 until 100) {
    println(gpuContains("\\101", Seq("A")).mkString(", "))
  }
}

For reference, here is the gpuContains method. I Added some println statements here.

private def gpuContains(cudfPattern: String, input: Seq[String]): Array[Boolean] = {
  println(cudfPattern)
  println(input.mkString(", "))
  val result = new Array[Boolean](input.length)
  withResource(ColumnVector.fromStrings(input: _*)) { cv =>
    withResource(cv.containsRe(cudfPattern)) { c =>
      withResource(c.copyToHost()) { hv =>
        result.indices.foreach(i => result(i) = hv.getBoolean(i))
      }
    }
  }
  result
}

Test output:

\101
A
false
\101
A
true
\101
A
true
...

If I remove the println statements from the gpuContains method then the result is always true.

Steps/Code to reproduce bug
See above.

Expected behavior
Result should be the same on every call.

Environment details (please complete the following information)
N/A

Additional context
N/A

[andygrove]

Related cuDF issue: rapidsai/cudf#9946

[andygrove]

For 22.02 we will fall back to CPU for regular expressions containing octal digits (this is already the case).

[andygrove]

We should go ahead and re-enable support octal digits (and the related tests) now that rapidsai/cudf#9993 was merged into cuDF