add warning message about using logreg in parallel (#2860)

Author: david-cortes-intel

doc/sources/algorithms.rst

@@ -92,7 +92,9 @@ Classification
        - ``class_weight`` != ``None``
        - Solver ``'newton-cg'`` with ``fit_intercept`` = ``False`` is not supported
      - Sparse data is not supported.
-     - Solver ``'newton-cg'`` is **only** available in :doc:`preview mode <preview>`.
+     - Solver ``'newton-cg'`` is **only** available in :doc:`preview mode <preview>`. **Important:** this estimator should not be used
+       in parallel Python threads - for concurrent fits (e.g. from :obj:`sklearn.model_selection.GridSearchCV`),
+       process-based parallelism should be used instead (default backend for :mod:`joblib`).
    * - :obj:`sklearn.linear_model.LogisticRegressionCV`
      - All parameters are supported except:
 

doc/sources/parallelism.rst

@@ -74,6 +74,13 @@ If ``n_jobs`` is not specified for underlying estimator(s), |sklearnex| sets it
     If concurrent calls are to be performed, process-based parallelism should
     be used instead.
 
+.. warning::
+    The patched classes :obj:`sklearn.linear_model.LogisticRegression` and
+    :obj:`sklearn.linear_model.LogisticRegressionCV` in particular are not
+    suitable for parallel calls in Python threads regardless of the ``n_jobs`` parameter.
+    Attempting to fit multiple logistic regression estimator objects in parallel
+    might result in crashes and incorrect estimations.
+
 Setting the `DEBUG` :ref:`verbosity setting <verbose>` will produce logs
 indicating when the number of threads used is different from the default
 (number of logical threads in the machine).