Change ON UPDATE SET NULL/SET DEFAULT referential actions to meet SQL spec.

Previously, when executing an ON UPDATE SET NULL or SET DEFAULT action for
a multicolumn MATCH SIMPLE foreign key constraint, we would set only those
referencing columns corresponding to referenced columns that were changed.
This is what the SQL92 standard said to do --- but more recent versions
of the standard say that all referencing columns should be set to null or
their default values, no matter exactly which referenced columns changed.
At least for SET DEFAULT, that is clearly saner behavior.  It's somewhat
debatable whether it's an improvement for SET NULL, but it appears that
other RDBMS systems read the spec this way.  So let's do it like that.

This is a release-notable behavioral change, although considering that
our documentation already implied it was done this way, the lack of
complaints suggests few people use such cases.

Author: tglsfdc

doc/src/sgml/ddl.sgml

@@ -735,7 +735,7 @@ CREATE TABLE t1 (
    </para>
 
    <para>
-    A table can contain more than one foreign key constraint.  This is
+    A table can have more than one foreign key constraint.  This is
     used to implement many-to-many relationships between tables.  Say
     you have tables about products and orders, but now you want to
     allow one order to contain possibly many products (which the
@@ -827,41 +827,52 @@ CREATE TABLE order_items (
     row(s) referencing it should be automatically deleted as well.
     There are two other options:
     <literal>SET NULL</literal> and <literal>SET DEFAULT</literal>.
-    These cause the referencing columns to be set to nulls or default
+    These cause the referencing column(s) in the referencing row(s)
+    to be set to nulls or their default
     values, respectively, when the referenced row is deleted.
     Note that these do not excuse you from observing any constraints.
     For example, if an action specifies <literal>SET DEFAULT</literal>
-    but the default value would not satisfy the foreign key, the
+    but the default value would not satisfy the foreign key constraint, the
     operation will fail.
    </para>
 
    <para>
     Analogous to <literal>ON DELETE</literal> there is also
     <literal>ON UPDATE</literal> which is invoked when a referenced
     column is changed (updated).  The possible actions are the same.
+    In this case, <literal>CASCADE</> means that the updated values of the
+    referenced column(s) should be copied into the referencing row(s).
    </para>
 
    <para>
+    Normally, a referencing row need not satisfy the foreign key constraint
+    if any of its referencing columns are null.  If <literal>MATCH FULL</>
+    is added to the foreign key declaration, a referencing row escapes
+    satisfying the constraint only if all its referencing columns are null
+    (so a mix of null and non-null values is guaranteed to fail a
+    <literal>MATCH FULL</> constraint).  If you don't want referencing rows
+    to be able to avoid satisfying the foreign key constraint, declare the
+    referencing column(s) as <literal>NOT NULL</>.
+   </para>
+
+   <para>
+    A foreign key must reference columns that either are a primary key or
+    form a unique constraint.  This means that the referenced columns always
+    have an index (the one underlying the primary key or unique constraint);
+    so checks on whether a referencing row has a match will be efficient.
     Since a <command>DELETE</command> of a row from the referenced table
     or an <command>UPDATE</command> of a referenced column will require
     a scan of the referencing table for rows matching the old value, it
-    is often a good idea to index the referencing columns.  Because this
+    is often a good idea to index the referencing columns too.  Because this
     is not always needed, and there are many choices available on how
     to index, declaration of a foreign key constraint does not
     automatically create an index on the referencing columns.
    </para>
 
    <para>
     More information about updating and deleting data is in <xref
-    linkend="dml">.
-   </para>
-
-   <para>
-    Finally, we should mention that a foreign key must reference
-    columns that either are a primary key or form a unique constraint.
-    If the foreign key references a unique constraint, there are some
-    additional possibilities regarding how null values are matched.
-    These are explained in the reference documentation for
+    linkend="dml">.  Also see the description of foreign key constraint
+    syntax in the reference documentation for
     <xref linkend="sql-createtable">.
    </para>
   </sect2>

doc/src/sgml/ref/create_table.sgml

@@ -585,8 +585,8 @@ CREATE [ [ GLOBAL | LOCAL ] { TEMPORARY | TEMP } | UNLOGGED ] TABLE [ IF NOT EXI
       These clauses specify a foreign key constraint, which requires
       that a group of one or more columns of the new table must only
       contain values that match values in the referenced
-      column(s) of some row of the referenced table.  If <replaceable
-      class="parameter">refcolumn</replaceable> is omitted, the
+      column(s) of some row of the referenced table.  If the <replaceable
+      class="parameter">refcolumn</replaceable> list is omitted, the
       primary key of the <replaceable class="parameter">reftable</replaceable>
       is used.  The referenced columns must be the columns of a non-deferrable
       unique or primary key constraint in the referenced table.  Note that
@@ -599,12 +599,16 @@ CREATE [ [ GLOBAL | LOCAL ] { TEMPORARY | TEMP } | UNLOGGED ] TABLE [ IF NOT EXI
       values of the referenced table and referenced columns using the
       given match type.  There are three match types: <literal>MATCH
       FULL</>, <literal>MATCH PARTIAL</>, and <literal>MATCH
-      SIMPLE</literal>, which is also the default.  <literal>MATCH
+      SIMPLE</literal> (which is the default).  <literal>MATCH
       FULL</> will not allow one column of a multicolumn foreign key
-      to be null unless all foreign key columns are null.
-      <literal>MATCH SIMPLE</literal> allows some foreign key columns
-      to be null while other parts of the foreign key are not
-      null. <literal>MATCH PARTIAL</> is not yet implemented.
+      to be null unless all foreign key columns are null; if they are all
+      null, the row is not required to have a match in the referenced table.
+      <literal>MATCH SIMPLE</literal> allows any of the foreign key columns
+      to be null; if any of them are null, the row is not required to have a
+      match in the referenced table.
+      <literal>MATCH PARTIAL</> is not yet implemented.
+      (Of course, <literal>NOT NULL</> constraints can be applied to the
+      referencing column(s) to prevent these cases from arising.)
      </para>
 
      <para>
@@ -652,8 +656,8 @@ CREATE [ [ GLOBAL | LOCAL ] { TEMPORARY | TEMP } | UNLOGGED ] TABLE [ IF NOT EXI
         <listitem>
          <para>
           Delete any rows referencing the deleted row, or update the
-          value of the referencing column to the new value of the
-          referenced column, respectively.
+          values of the referencing column(s) to the new values of the
+          referenced columns, respectively.
          </para>
         </listitem>
        </varlistentry>
@@ -672,6 +676,8 @@ CREATE [ [ GLOBAL | LOCAL ] { TEMPORARY | TEMP } | UNLOGGED ] TABLE [ IF NOT EXI
         <listitem>
          <para>
           Set the referencing column(s) to their default values.
+          (There must be a row in the referenced table matching the default
+          values, if they are not null, or the operation will fail.)
          </para>
         </listitem>
        </varlistentry>
@@ -680,8 +686,8 @@ CREATE [ [ GLOBAL | LOCAL ] { TEMPORARY | TEMP } | UNLOGGED ] TABLE [ IF NOT EXI
 
      <para>
       If the referenced column(s) are changed frequently, it might be wise to
-      add an index to the foreign key column so that referential actions
-      associated with the foreign key column can be performed more
+      add an index to the referencing column(s) so that referential actions
+      associated with the foreign key constraint can be performed more
       efficiently.
      </para>
     </listitem>

src/backend/utils/adt/ri_triggers.c

@@ -207,11 +207,6 @@ static void ri_BuildQueryKeyPkCheck(RI_QueryKey *key,
 						int32 constr_queryno);
 static bool ri_KeysEqual(Relation rel, HeapTuple oldtup, HeapTuple newtup,
 			 const RI_ConstraintInfo *riinfo, bool rel_is_pk);
-static bool ri_AllKeysUnequal(Relation rel, HeapTuple oldtup, HeapTuple newtup,
-				  const RI_ConstraintInfo *riinfo, bool rel_is_pk);
-static bool ri_OneKeyEqual(Relation rel, int column,
-			   HeapTuple oldtup, HeapTuple newtup,
-			   const RI_ConstraintInfo *riinfo, bool rel_is_pk);
 static bool ri_AttributesEqual(Oid eq_opr, Oid typeid,
 				   Datum oldvalue, Datum newvalue);
 static bool ri_Check_Pk_Match(Relation pk_rel, Relation fk_rel,
@@ -1950,7 +1945,6 @@ RI_FKey_setnull_upd(PG_FUNCTION_ARGS)
 	RI_QueryKey qkey;
 	SPIPlanPtr	qplan;
 	int			i;
-	bool		use_cached_query;
 
 	/*
 	 * Check that this is a valid trigger call on the right time and event.
@@ -1985,7 +1979,7 @@ RI_FKey_setnull_upd(PG_FUNCTION_ARGS)
 			/* ----------
 			 * SQL3 11.9 <referential constraint definition>
 			 *	General rules 7) a) ii) 2):
-			 *		MATCH FULL
+			 *		MATCH SIMPLE/FULL
 			 *			... ON UPDATE SET NULL
 			 * ----------
 			 */
@@ -2026,29 +2020,10 @@ RI_FKey_setnull_upd(PG_FUNCTION_ARGS)
 			if (SPI_connect() != SPI_OK_CONNECT)
 				elog(ERROR, "SPI_connect failed");
 
-			/*
-			 * "MATCH SIMPLE" only changes columns corresponding to the
-			 * referenced columns that have changed in pk_rel.	This means the
-			 * "SET attrn=NULL [, attrn=NULL]" string will be change as well.
-			 * In this case, we need to build a temporary plan rather than use
-			 * our cached plan, unless the update happens to change all
-			 * columns in the key.	Fortunately, for the most common case of a
-			 * single-column foreign key, this will be true.
-			 *
-			 * In case you're wondering, the inequality check works because we
-			 * know that the old key value has no NULLs (see above).
-			 */
-
-			use_cached_query = (riinfo.confmatchtype == FKCONSTR_MATCH_FULL) ||
-				ri_AllKeysUnequal(pk_rel, old_row, new_row,
-								  &riinfo, true);
-
 			/*
 			 * Fetch or prepare a saved plan for the set null update operation
-			 * if possible, or build a temporary plan if not.
 			 */
-			if (!use_cached_query ||
-				(qplan = ri_FetchPreparedPlan(&qkey)) == NULL)
+			if ((qplan = ri_FetchPreparedPlan(&qkey)) == NULL)
 			{
 				StringInfoData querybuf;
 				StringInfoData qualbuf;
@@ -2080,37 +2055,23 @@ RI_FKey_setnull_upd(PG_FUNCTION_ARGS)
 
 					quoteOneName(attname,
 								 RIAttName(fk_rel, riinfo.fk_attnums[i]));
-
-					/*
-					 * MATCH SIMPLE - only change columns corresponding
-					 * to changed columns in pk_rel's key
-					 */
-					if (riinfo.confmatchtype == FKCONSTR_MATCH_FULL ||
-						!ri_OneKeyEqual(pk_rel, i, old_row, new_row,
-										&riinfo, true))
-					{
-						appendStringInfo(&querybuf,
-										 "%s %s = NULL",
-										 querysep, attname);
-						querysep = ",";
-					}
+					appendStringInfo(&querybuf,
+									 "%s %s = NULL",
+									 querysep, attname);
 					sprintf(paramname, "$%d", i + 1);
 					ri_GenerateQual(&qualbuf, qualsep,
 									paramname, pk_type,
 									riinfo.pf_eq_oprs[i],
 									attname, fk_type);
+					querysep = ",";
 					qualsep = "AND";
 					queryoids[i] = pk_type;
 				}
 				appendStringInfoString(&querybuf, qualbuf.data);
 
-				/*
-				 * Prepare the plan.  Save it only if we're building the
-				 * "standard" plan.
-				 */
+				/* Prepare and save the plan */
 				qplan = ri_PlanCheck(querybuf.data, riinfo.nkeys, queryoids,
-									 &qkey, fk_rel, pk_rel,
-									 use_cached_query);
+									 &qkey, fk_rel, pk_rel, true);
 			}
 
 			/*
@@ -2463,25 +2424,15 @@ RI_FKey_setdefault_upd(PG_FUNCTION_ARGS)
 
 					quoteOneName(attname,
 								 RIAttName(fk_rel, riinfo.fk_attnums[i]));
-
-					/*
-					 * MATCH SIMPLE - only change columns corresponding
-					 * to changed columns in pk_rel's key
-					 */
-					if (riinfo.confmatchtype == FKCONSTR_MATCH_FULL ||
-						!ri_OneKeyEqual(pk_rel, i, old_row, new_row,
-										&riinfo, true))
-					{
-						appendStringInfo(&querybuf,
-										 "%s %s = DEFAULT",
-										 querysep, attname);
-						querysep = ",";
-					}
+					appendStringInfo(&querybuf,
+									 "%s %s = DEFAULT",
+									 querysep, attname);
 					sprintf(paramname, "$%d", i + 1);
 					ri_GenerateQual(&qualbuf, qualsep,
 									paramname, pk_type,
 									riinfo.pf_eq_oprs[i],
 									attname, fk_type);
+					querysep = ",";
 					qualsep = "AND";
 					queryoids[i] = pk_type;
 				}
@@ -3857,120 +3808,6 @@ ri_KeysEqual(Relation rel, HeapTuple oldtup, HeapTuple newtup,
 }
 
 
-/* ----------
- * ri_AllKeysUnequal -
- *
- *	Check if all key values in OLD and NEW are not equal.
- * ----------
- */
-static bool
-ri_AllKeysUnequal(Relation rel, HeapTuple oldtup, HeapTuple newtup,
-				  const RI_ConstraintInfo *riinfo, bool rel_is_pk)
-{
-	TupleDesc	tupdesc = RelationGetDescr(rel);
-	const int16 *attnums;
-	const Oid  *eq_oprs;
-	int			i;
-
-	if (rel_is_pk)
-	{
-		attnums = riinfo->pk_attnums;
-		eq_oprs = riinfo->pp_eq_oprs;
-	}
-	else
-	{
-		attnums = riinfo->fk_attnums;
-		eq_oprs = riinfo->ff_eq_oprs;
-	}
-
-	for (i = 0; i < riinfo->nkeys; i++)
-	{
-		Datum		oldvalue;
-		Datum		newvalue;
-		bool		isnull;
-
-		/*
-		 * Get one attribute's oldvalue. If it is NULL - they're not equal.
-		 */
-		oldvalue = SPI_getbinval(oldtup, tupdesc, attnums[i], &isnull);
-		if (isnull)
-			continue;
-
-		/*
-		 * Get one attribute's newvalue. If it is NULL - they're not equal.
-		 */
-		newvalue = SPI_getbinval(newtup, tupdesc, attnums[i], &isnull);
-		if (isnull)
-			continue;
-
-		/*
-		 * Compare them with the appropriate equality operator.
-		 */
-		if (ri_AttributesEqual(eq_oprs[i], RIAttType(rel, attnums[i]),
-							   oldvalue, newvalue))
-			return false;		/* found two equal items */
-	}
-
-	return true;
-}
-
-
-/* ----------
- * ri_OneKeyEqual -
- *
- *	Check if one key value in OLD and NEW is equal.  Note column is indexed
- *	from zero.
- *
- *	ri_KeysEqual could call this but would run a bit slower.  For
- *	now, let's duplicate the code.
- * ----------
- */
-static bool
-ri_OneKeyEqual(Relation rel, int column, HeapTuple oldtup, HeapTuple newtup,
-			   const RI_ConstraintInfo *riinfo, bool rel_is_pk)
-{
-	TupleDesc	tupdesc = RelationGetDescr(rel);
-	const int16 *attnums;
-	const Oid  *eq_oprs;
-	Datum		oldvalue;
-	Datum		newvalue;
-	bool		isnull;
-
-	if (rel_is_pk)
-	{
-		attnums = riinfo->pk_attnums;
-		eq_oprs = riinfo->pp_eq_oprs;
-	}
-	else
-	{
-		attnums = riinfo->fk_attnums;
-		eq_oprs = riinfo->ff_eq_oprs;
-	}
-
-	/*
-	 * Get one attribute's oldvalue. If it is NULL - they're not equal.
-	 */
-	oldvalue = SPI_getbinval(oldtup, tupdesc, attnums[column], &isnull);
-	if (isnull)
-		return false;
-
-	/*
-	 * Get one attribute's newvalue. If it is NULL - they're not equal.
-	 */
-	newvalue = SPI_getbinval(newtup, tupdesc, attnums[column], &isnull);
-	if (isnull)
-		return false;
-
-	/*
-	 * Compare them with the appropriate equality operator.
-	 */
-	if (!ri_AttributesEqual(eq_oprs[column], RIAttType(rel, attnums[column]),
-							oldvalue, newvalue))
-		return false;
-
-	return true;
-}
-
 /* ----------
  * ri_AttributesEqual -
  *