Fix tsmatchsel() to account properly for null rows.

ts_typanalyze.c computes MCE statistics as fractions of the non-null rows,
which seems fairly reasonable, and anyway changing it in released versions
wouldn't be a good idea.  But then ts_selfuncs.c has to account for that.
Failure to do so results in overestimates in columns with a significant
fraction of null documents.  Back-patch to 8.4 where this stuff was
introduced.

Jesper Krogh

diff --git a/src/backend/tsearch/ts_selfuncs.c b/src/backend/tsearch/ts_selfuncs.c
index e7194ce..b679b75 100644 (file)
--- a/src/backend/tsearch/ts_selfuncs.c
+++ b/src/backend/tsearch/ts_selfuncs.c
@@ -189,11 +189,17 @@ tsquerysel(VariableStatData *vardata, Datum constval)
                        /* No most-common-elements info, so do without */
                        selec = tsquery_opr_selec_no_stats(query);
                }
+
+               /*
+                * MCE stats count only non-null rows, so adjust for null rows.
+                */
+               selec *= (1.0 - stats->stanullfrac);
        }
        else
        {
                /* No stats at all, so do without */
                selec = tsquery_opr_selec_no_stats(query);
+               /* we assume no nulls here, so no stanullfrac correction */
        }
 
        return selec;

diff --git a/src/include/catalog/pg_statistic.h b/src/include/catalog/pg_statistic.h
index 7977743..f5965b5 100644 (file)
--- a/src/include/catalog/pg_statistic.h
+++ b/src/include/catalog/pg_statistic.h
@@ -246,6 +246,8 @@ typedef FormData_pg_statistic *Form_pg_statistic;
  * type with identifiable elements (for instance, tsvector).  staop contains
  * the equality operator appropriate to the element type.  stavalues contains
  * the most common element values, and stanumbers their frequencies.  Unlike
+ * MCV slots, frequencies are measured as the fraction of non-null rows the
+ * element value appears in, not the frequency of all rows.  Also unlike
  * MCV slots, the values are sorted into order (to support binary search
  * for a particular value).  Since this puts the minimum and maximum
  * frequencies at unpredictable spots in stanumbers, there are two extra