Improve hash join's handling of tuples with null join keys.

In a plain join, we can just summarily discard an input tuple
with null join key(s), since it cannot match anything from
the other side of the join (assuming a strict join operator).
However, if the tuple comes from the outer side of an outer join
then we have to emit it with null-extension of the other side.

Up to now, hash joins did that by inserting the tuple into the hash
table as though it were a normal tuple.  This is unnecessarily
inefficient though, since the required processing is far simpler than
for a potentially-matchable tuple.  Worse, if there are a lot of such
tuples they will bloat the hash bucket they go into, possibly causing
useless repeated attempts to split that bucket or increase the number
of batches.  We have a report of a large join vainly creating many
thousands of batches when faced with such input.

This patch improves the situation by keeping such tuples out of the
hash table altogether, instead pushing them into a separate tuplestore
from which we return them later.  (One might consider trying to return
them immediately; but that would require substantial refactoring, and
it doesn't work anyway for the case where we rescan an unmodified hash
table.)  This works even in parallel hash joins, because whichever
worker reads a null-keyed tuple can just return it; there's no need
for consultation with other workers.  Thus the tuplestores are local
storage even in a parallel join.

Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/3061845.1746486714@sss.pgh.pa.us

Author: tglsfdc

src/backend/executor/execExpr.c

@@ -4274,25 +4274,27 @@ ExecBuildHash32FromAttrs(TupleDesc desc, const TupleTableSlotOps *ops,
  * 'hash_exprs'.  When multiple expressions are present, the hash values
  * returned by each hash function are combined to produce a single hash value.
  *
+ * If any hash_expr yields NULL and the corresponding hash operator is strict,
+ * the created ExprState will return NULL.  (If the operator is not strict,
+ * we treat NULL values as having a hash value of zero.  The hash functions
+ * themselves are always treated as strict.)
+ *
  * desc: tuple descriptor for the to-be-hashed expressions
  * ops: TupleTableSlotOps for the TupleDesc
  * hashfunc_oids: Oid for each hash function to call, one for each 'hash_expr'
- * collations: collation to use when calling the hash function.
- * hash_expr: list of expressions to hash the value of
- * opstrict: array corresponding to the 'hashfunc_oids' to store op_strict()
+ * collations: collation to use when calling the hash function
+ * hash_exprs: list of expressions to hash the value of
+ * opstrict: strictness flag for each hash function's comparison operator
  * parent: PlanState node that the 'hash_exprs' will be evaluated at
  * init_value: Normally 0, but can be set to other values to seed the hash
  * with some other value.  Using non-zero is slightly less efficient but can
  * be useful.
- * keep_nulls: if true, evaluation of the returned ExprState will abort early
- * returning NULL if the given hash function is strict and the Datum to hash
- * is null.  When set to false, any NULL input Datums are skipped.
  */
 ExprState *
 ExecBuildHash32Expr(TupleDesc desc, const TupleTableSlotOps *ops,
 					const Oid *hashfunc_oids, const List *collations,
 					const List *hash_exprs, const bool *opstrict,
-					PlanState *parent, uint32 init_value, bool keep_nulls)
+					PlanState *parent, uint32 init_value)
 {
 	ExprState  *state = makeNode(ExprState);
 	ExprEvalStep scratch = {0};
@@ -4369,8 +4371,8 @@ ExecBuildHash32Expr(TupleDesc desc, const TupleTableSlotOps *ops,
 		fmgr_info(funcid, finfo);
 
 		/*
-		 * Build the steps to evaluate the hash function's argument have it so
-		 * the value of that is stored in the 0th argument of the hash func.
+		 * Build the steps to evaluate the hash function's argument, placing
+		 * the value in the 0th argument of the hash func.
 		 */
 		ExecInitExprRec(expr,
 						state,
@@ -4405,7 +4407,7 @@ ExecBuildHash32Expr(TupleDesc desc, const TupleTableSlotOps *ops,
 		scratch.d.hashdatum.fcinfo_data = fcinfo;
 		scratch.d.hashdatum.fn_addr = finfo->fn_addr;
 
-		scratch.opcode = opstrict[i] && !keep_nulls ? strict_opcode : opcode;
+		scratch.opcode = opstrict[i] ? strict_opcode : opcode;
 		scratch.d.hashdatum.jumpdone = -1;
 
 		ExprEvalPushStep(state, &scratch);

src/backend/executor/nodeHash.c

@@ -153,8 +153,11 @@ MultiExecPrivateHash(HashState *node)
 	econtext = node->ps.ps_ExprContext;
 
 	/*
-	 * Get all tuples from the node below the Hash node and insert into the
-	 * hash table (or temp files).
+	 * Get all tuples from the node below the Hash node and insert the
+	 * potentially-matchable ones into the hash table (or temp files).  Tuples
+	 * that can't possibly match because they have null join keys are dumped
+	 * into a separate tuplestore, or just summarily discarded if we don't
+	 * need to emit them with null-extension.
 	 */
 	for (;;)
 	{
@@ -174,6 +177,7 @@ MultiExecPrivateHash(HashState *node)
 
 		if (!isnull)
 		{
+			/* normal case with a non-null join key */
 			uint32		hashvalue = DatumGetUInt32(hashdatum);
 			int			bucketNumber;
 
@@ -192,6 +196,14 @@ MultiExecPrivateHash(HashState *node)
 			}
 			hashtable->totalTuples += 1;
 		}
+		else if (node->keep_null_tuples)
+		{
+			/* null join key, but we must save tuple to be emitted later */
+			if (node->null_tuple_store == NULL)
+				node->null_tuple_store = ExecHashBuildNullTupleStore(hashtable);
+			tuplestore_puttupleslot(node->null_tuple_store, slot);
+		}
+		/* else we can discard the tuple immediately */
 	}
 
 	/* resize the hash table if needed (NTUP_PER_BUCKET exceeded) */
@@ -222,7 +234,6 @@ MultiExecParallelHash(HashState *node)
 	HashJoinTable hashtable;
 	TupleTableSlot *slot;
 	ExprContext *econtext;
-	uint32		hashvalue;
 	Barrier    *build_barrier;
 	int			i;
 
@@ -282,6 +293,7 @@ MultiExecParallelHash(HashState *node)
 			for (;;)
 			{
 				bool		isnull;
+				uint32		hashvalue;
 
 				slot = ExecProcNode(outerNode);
 				if (TupIsNull(slot))
@@ -295,8 +307,19 @@ MultiExecParallelHash(HashState *node)
 																	 &isnull));
 
 				if (!isnull)
+				{
+					/* normal case with a non-null join key */
 					ExecParallelHashTableInsert(hashtable, slot, hashvalue);
-				hashtable->partialTuples++;
+					hashtable->partialTuples++;
+				}
+				else if (node->keep_null_tuples)
+				{
+					/* null join key, but save tuple to be emitted later */
+					if (node->null_tuple_store == NULL)
+						node->null_tuple_store = ExecHashBuildNullTupleStore(hashtable);
+					tuplestore_puttupleslot(node->null_tuple_store, slot);
+				}
+				/* else we can discard the tuple immediately */
 			}
 
 			/*
@@ -404,14 +427,10 @@ ExecInitHash(Hash *node, EState *estate, int eflags)
 
 	Assert(node->plan.qual == NIL);
 
-	/*
-	 * Delay initialization of hash_expr until ExecInitHashJoin().  We cannot
-	 * build the ExprState here as we don't yet know the join type we're going
-	 * to be hashing values for and we need to know that before calling
-	 * ExecBuildHash32Expr as the keep_nulls parameter depends on the join
-	 * type.
-	 */
+	/* these fields will be filled by ExecInitHashJoin() */
 	hashstate->hash_expr = NULL;
+	hashstate->null_tuple_store = NULL;
+	hashstate->keep_null_tuples = false;
 
 	return hashstate;
 }
@@ -2753,6 +2772,31 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 	}
 }
 
+/*
+ * Build a tuplestore suitable for holding null-keyed input tuples.
+ * (This function doesn't care whether it's for outer or inner tuples.)
+ *
+ * Note that in a parallel hash join, each worker has its own tuplestore(s)
+ * for these.  There's no need to interact with other workers to decide
+ * what to do with them.  So they're always in private storage.
+ */
+Tuplestorestate *
+ExecHashBuildNullTupleStore(HashJoinTable hashtable)
+{
+	Tuplestorestate *tstore;
+	MemoryContext oldcxt;
+
+	/*
+	 * We keep the tuplestore in the hashCxt to ensure it won't go away too
+	 * soon.  Size it at work_mem/16 so that it doesn't bloat the node's space
+	 * consumption too much.
+	 */
+	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	tstore = tuplestore_begin_heap(false, false, work_mem / 16);
+	MemoryContextSwitchTo(oldcxt);
+	return tstore;
+}
+
 /*
  * Reserve space in the DSM segment for instrumentation data.
  */