[SPARK-20214] pyspark linalg _convert_to_vector should check for sorted indices

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.5.2, 1.6.3, 2.0.2, 2.1.1, 2.2.0
• Fix Version/s: 2.0.3, 2.1.1, 2.2.0
• Component/s: ML, MLlib, PySpark, Tests
• Target Version/s: 2.0.3, 2.1.2, 2.2.0
• Assignee: L. C. Hsieh
• Reporter: Joseph K. Bradley
• Shepherd: Joseph K. Bradley
• Created: 04/Apr/17 22:03
• Updated: 06/Apr/17 00:59
• Resolved: 06/Apr/17 00:51

Description

I've seen a few failures of this line: https://github.com/apache/spark/blame/402bf2a50ddd4039ff9f376b641bd18fffa54171/python/pyspark/mllib/tests.py#L847

It converts a scipy.sparse.lil_matrix to a dok_matrix and then to a pyspark.mllib.linalg.Vector. The failure happens in the conversion to a vector and indicates that the dok_matrix is not returning its values in sorted order. (Actually, the failure is in _convert_to_vector, which converts the dok_matrix to a csc_matrix and then passes the CSC data to the MLlib Vector constructor.) Here's the stack trace:

Traceback (most recent call last):
  File "/home/jenkins/workspace/python/pyspark/mllib/tests.py", line 847, in test_serialize
    self.assertEqual(sv, _convert_to_vector(lil.todok()))
  File "/home/jenkins/workspace/python/pyspark/mllib/linalg/__init__.py", line 78, in _convert_to_vector
    return SparseVector(l.shape[0], csc.indices, csc.data)
  File "/home/jenkins/workspace/python/pyspark/mllib/linalg/__init__.py", line 556, in __init__
    % (self.indices[i], self.indices[i + 1]))
TypeError: Indices 3 and 1 are not strictly increasing

This seems like a bug in _convert_to_vector, where we really should check csc_matrix.has_sorted_indices first.

I haven't seen this bug in pyspark.ml.linalg, but it probably exists there too.

Issue Links

• links to: [Github] Pull Request #17532 (viirya)