Linregr: Fix performance regression for heteroskedasticity computation

Pivotal Tracker: 74271378
JIRA: MADLIB-863

Details:
    Linear regression aggregate returns a composite type that is directly
    joined to the source table to compute heteroskedasticity. The only
    value needed from the model is the coefficient, but the composite
    type leads the planner to use the whole result during the join. This
    led to a performance regression when the result type included the
    covariance matrix in v1.6. This caused a join with an extremely wide
    row even though only the coefficient was needed. This commit fixes
    the problem by expanding the composite type to its individual members
    in the model table.

src/ports/postgres/modules/regress/linear.py_in

@@ -34,9 +34,9 @@ def linregr_train(schema_madlib, source_table, out_table,
             CREATE TEMP TABLE {temp_lin_rst} AS
             SELECT
                 {group_str_sel}
-                {schema_madlib}.linregr(
+                ({schema_madlib}.linregr(
                     {dependent_varname},
-                    {independent_varname}) AS lin_rst,
+                    {independent_varname})).*,
                 count(*) AS num_rows
             FROM
                 {source_table}
@@ -61,7 +61,7 @@ def linregr_train(schema_madlib, source_table, out_table,
                     {schema_madlib}.heteroskedasticity_test_linregr(
                         {dependent_varname},
                         {independent_varname},
-                        (lin_rst).coef) AS hsk_rst
+                        {temp_lin_rst}.coef) AS hsk_rst
                 FROM
                     {source_table} {join_str} {temp_lin_rst} {using_str}
                 {group_str}
@@ -90,23 +90,23 @@ def linregr_train(schema_madlib, source_table, out_table,
             CREATE TABLE {out_table} AS
             SELECT
                 {group_str_sel}
-                (lin.lin_rst).coef,
-                (lin.lin_rst).r2,
-                (lin.lin_rst).std_err,
-                (lin.lin_rst).t_stats,
-                (lin.lin_rst).p_values,
-                (lin.lin_rst).condition_no,
+                coef,
+                r2,
+                std_err,
+                t_stats,
+                p_values,
+                condition_no,
                 {bp_stats}
                 {bp_p_value}
-                CASE WHEN (lin.lin_rst).num_processed IS NULL
+                CASE WHEN num_processed IS NULL
                     THEN 0
-                    ELSE (lin.lin_rst).num_processed
+                    ELSE num_processed
                 END AS num_rows_processed,
-                CASE WHEN (lin.lin_rst).num_processed IS NULL
+                CASE WHEN num_processed IS NULL
                     THEN lin.num_rows
-                    ELSE lin.num_rows - (lin.lin_rst).num_processed
+                    ELSE lin.num_rows - num_processed
                 END AS num_missing_rows_skipped,
-                (lin.lin_rst).vcov as variance_covariance
+                vcov as variance_covariance
             FROM
                 {temp_lin_rst} AS lin {join_str} {using_str}
            """.format(out_table=out_table, group_str_sel=group_str_sel,
@@ -176,12 +176,12 @@ def _validate_args(schema_madlib, source_table, out_table, dependent_varname,
     # an informative error messages when they are not valid
 
     _assert(heteroskedasticity_option is not None and
-        heteroskedasticity_option in (True, False),
-        "Linregr error: Invalid heteroskedasticity_option")
+            heteroskedasticity_option in (True, False),
+            "Linregr error: Invalid heteroskedasticity_option")
 
     if grouping_cols is not None:
         _assert(grouping_cols != '',
-            "Linregr error: Invalid grouping columns name!")
+                "Linregr error: Invalid grouping columns name!")
         grouping_list = _string_to_array(grouping_cols)
         _assert(columns_exist_in_table(
             source_table, grouping_list, schema_madlib),
@@ -194,10 +194,10 @@ def _validate_args(schema_madlib, source_table, out_table, dependent_varname,
         if heteroskedasticity_option is not None:
             predefined.update(frozenset(('bp_stats', 'bp_p_value')))
         intersect = frozenset(grouping_list).intersection(predefined)
-        _assert(len(intersect) == 0,
-            "Linregr error: Conflicted grouping column name.\n"
-            "Predefined name(s) {0} are not allow!".format(
-            ', '.join(intersect)))
+        _assert(not intersect,
+                "Linregr error: Conflicting grouping column name.\n"
+                "Predefined name(s) {0} are not allowed!".
+                format(', '.join(intersect)))
 
 # ------------------------------------------------------------------------------
 # -- Online help function ------------------------------------------------------