[SPARK-18341] Eliminate use of SingularMatrixException in WeightedLeastSquares logic

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Won't Fix
• Component/s: ML
• Assignee: Unassigned
• Reporter: Joseph K. Bradley
• Created: 08/Nov/16 03:32
• Updated: 06/Feb/17 11:05
• Resolved: 06/Feb/17 11:05

Description

WeightedLeastSquares uses an Exception to implement fallback logic for which solver to use: https://github.com/apache/spark/blob/6f3697136aa68dc39d3ce42f43a7af554d2a3bf9/mllib/src/main/scala/org/apache/spark/ml/optim/WeightedLeastSquares.scala#L258

We should use an error code instead of an exception.

• Note the error code should be internal, not a public API.
• We may be able to eliminate the SingularMatrixException class.

Issue Links

• is related to: SPARK-17748 One-pass algorithm for linear regression with L1 and elastic-net penalties
• links to: [Github] Pull Request #16002 (yanboliang)