Details
-
Bug
-
Status: Resolved
-
Critical
-
Resolution: Fixed
-
1.1.0
-
None
Description
There is a breaking bug in PySpark's sampling methods when run with NumPy v1.9. This is the version of NumPy included with the current Anaconda distribution (v2.1); this is a popular distribution, and is likely to affect many users.
Steps to reproduce are:
foo = sc.parallelize(range(1000),5) foo.takeSample(False, 10)
Returns:
PythonException: Traceback (most recent call last): File "/Users/freemanj11/code/spark-1.1.0-bin-hadoop1/python/pyspark/worker.py", line 79, in main serializer.dump_stream(func(split_index, iterator), outfile) File "/Users/freemanj11/code/spark-1.1.0-bin-hadoop1/python/pyspark/serializers.py", line 196, in dump_stream self.serializer.dump_stream(self._batched(iterator), stream) File "/Users/freemanj11/code/spark-1.1.0-bin-hadoop1/python/pyspark/serializers.py", line 127, in dump_stream for obj in iterator: File "/Users/freemanj11/code/spark-1.1.0-bin-hadoop1/python/pyspark/serializers.py", line 185, in _batched for item in iterator: File "/Users/freemanj11/code/spark-1.1.0-bin-hadoop1/python/pyspark/rddsampler.py", line 116, in func if self.getUniformSample(split) <= self._fraction: File "/Users/freemanj11/code/spark-1.1.0-bin-hadoop1/python/pyspark/rddsampler.py", line 58, in getUniformSample self.initRandomGenerator(split) File "/Users/freemanj11/code/spark-1.1.0-bin-hadoop1/python/pyspark/rddsampler.py", line 44, in initRandomGenerator self._random = numpy.random.RandomState(self._seed) File "mtrand.pyx", line 610, in mtrand.RandomState.__init__ (numpy/random/mtrand/mtrand.c:7397) File "mtrand.pyx", line 646, in mtrand.RandomState.seed (numpy/random/mtrand/mtrand.c:7697) ValueError: Seed must be between 0 and 4294967295
In PySpark's RDDSamplerBase class from pyspark.rddsampler we use:
self._seed = seed if seed is not None else random.randint(0, sys.maxint)
In previous versions of NumPy a random seed larger than 2 ** 32 would silently get truncated to 2 ** 32. This was fixed in a recent patch (https://github.com/numpy/numpy/commit/6b1a1205eac6fe5d162f16155d500765e8bca53c). But sampling (0, sys.maxint) often yields ints larger than 2 ** 32, which effectively breaks sampling operations in PySpark (unless the seed is set manually).
I am putting a PR together now (the fix is very simple!).