When computing STAT1 values using ANALYZE, if a ratio comes out to be between

1.0 and 1.1, then round it down to 1 rather than the using the default rounding
rule of changing it to 2.  The reduces the estimation error for the
case where a column value is very nearly, but not quite unique.

FossilOrigin-Name: eb59c46a5aed69bc6fd096997bf24c082e533c1085439f6ec1fbe5ff78e8b374

1 files changed, 7 insertions, 1 deletions

diff --git a/src/analyze.c b/src/analyze.c
index e8699653f..39009899a 100644
--- a/src/analyze.c
+++ b/src/analyze.c
@@ -847,9 +847,14 @@ static void statGet(
     **   * "WHERE a=? AND b=?" matches 2 rows.
     **
     ** If D is the count of distinct values and K is the total number of 
-    ** rows, then each estimate is computed as:
+    ** rows, then each estimate is usually computed as:
     **
     **        I = (K+D-1)/D
+    **
+    ** In other words, I is K/D rounded up to the next whole integer.
+    ** However, if I is between 1.0 and 1.1 (in other words if I is
+    ** close to 1.0 but just a little larger) then do not round up but
+    ** instead keep the I value at 1.0.
     */
     sqlite3_str sStat;   /* Text of the constructed "stat" line */
     int i;               /* Loop counter */
@@ -860,6 +865,7 @@ static void statGet(
     for(i=0; i<p->nKeyCol; i++){
       u64 nDistinct = p->current.anDLt[i] + 1;
       u64 iVal = (p->nRow + nDistinct - 1) / nDistinct;
+      if( iVal==2 && p->nRow*10 <= nDistinct*11 ) iVal = 1;
       sqlite3_str_appendf(&sStat, " %llu", iVal);
       assert( p->current.anEq[i] );
     }