Parallel DEDUP_SEMI and DEDUP_SEMI_REVERSE Join.(A new approach to pr…

…ocess Semi Join Query in Parallel of MPP) (#653)

To handle semi join in parallel, we have enabled Parallel Semi join.
This commit enable DEDUP_SEMI and DEDUP_SEMI_REVERSE join in parallel
to handle semi join cases which could be a win in MPP mode.

Both parallel-oblivious and parallel-aware are enabled, for an example
of parallel-aware case:

select * from foo where exists (select 1 from bar where foo.a = bar.b);
                                QUERY PLAN
---------------------------------------------------------------------------
 Gather Motion 6:1  (slice1; segments: 6)
   ->  HashAggregate
         Group Key: (RowIdExpr)
         ->  Redistribute Motion 6:6  (slice2; segments: 6)
               Hash Key: (RowIdExpr)
               Hash Module: 3
               ->  Parallel Hash Join
                     Hash Cond: (bar.b = foo.a)
                     ->  Parallel Seq Scan on bar
                     ->  Parallel Hash
                           ->  Broadcast Workers Motion 6:6  (slice3;
segments: 6)
                                 ->  Parallel Seq Scan on foo
 Optimizer: Postgres query optimizer
(13 rows)

For DEDUP_SEMI or DEDUP_SEMI_REVERSE join, each process need a unique
RowIdExpr to identify unique rows, which is assigned with a baseline
when building paths.

It's ok for non-parallel plan, but in parallel mode there are multiple
processes on same segment, RowIdExpr in not unique then.
To enable that, add ParallelWorkerNumberOfSlice to identify worker id
of a parallel plan of a slice.
When rowidexpr is used, it's executed by 48 bits and left other 16 bits
for segment_id. In parallel mode, we have to make more room for parallel
worker id within segment_id's bits. This is done during planner with
checks, in case that there are many segments with many parallel workers
(which rarely happens).

Authored-by: Zhang Mingli [email protected]

src/backend/access/transam/parallel.c

@@ -142,6 +142,8 @@ typedef struct FixedParallelState
  * worker will get a different parallel worker number.
  */
 int			ParallelWorkerNumber = -1;
+int			ParallelWorkerNumberOfSlice = -1;
+int			TotalParallelWorkerNumberOfSlice = 0;
 
 /* Is there a parallel message pending which we need to receive? */
 volatile bool ParallelMessagePending = false;
@@ -1716,6 +1718,8 @@ void GpDestroyParallelDSMEntry()
 		ParallelSession->area = NULL;
 	}
 	LWLockRelease(GpParallelDSMHashLock);
+	ParallelWorkerNumberOfSlice = -1;
+	TotalParallelWorkerNumberOfSlice = 0;
 }
 
 void
@@ -1805,6 +1809,8 @@ GpInsertParallelDSMHash(PlanState *planstate)
 		entry->tolaunch = parallel_workers - 1;
 		entry->parallel_workers = parallel_workers;
 		entry->temp_worker_id = 0;
+		ParallelWorkerNumberOfSlice = 0; /* The first worker. */
+		Assert(TotalParallelWorkerNumberOfSlice == parallel_workers);
 
 		/* Create a DSA area that can be used by the leader and all workers. */
 		char	   *area_space = shm_toc_allocate(entry->toc, dsa_minsize);
@@ -1864,7 +1870,7 @@ GpInsertParallelDSMHash(PlanState *planstate)
 			.nworkers = parallel_workers,
 			.worker_id = entry->temp_worker_id,
 		};
-
+		ParallelWorkerNumberOfSlice = ctx.worker_id;
 		InitializeGpParallelWorkers(planstate, &ctx);
 	}
 

src/backend/cdb/cdbpath.c

@@ -44,6 +44,7 @@
 #include "cdb/cdbutil.h"
 #include "cdb/cdbvars.h"
 
+#include "port/pg_bitutils.h"
 
 typedef struct
 {
@@ -2998,6 +2999,8 @@ cdbpath_motion_for_parallel_join(PlannerInfo *root,
 	CdbPathLocus_MakeNull(&inner.move_to);
 	outer.isouter = true;
 	inner.isouter = false;
+	int outerParallel = outer.locus.parallel_workers;
+	int innerParallel = inner.locus.parallel_workers;
 
 	Assert(cdbpathlocus_is_valid(outer.locus));
 	Assert(cdbpathlocus_is_valid(inner.locus));
@@ -3091,6 +3094,9 @@ cdbpath_motion_for_parallel_join(PlannerInfo *root,
 		case JOIN_INNER:
 			break;
 		case JOIN_SEMI:
+			if (!enable_parallel_semi_join)
+				goto fail;
+			/* FALLTHROUGH */
 		case JOIN_ANTI:
 		case JOIN_LEFT:
 		case JOIN_LASJ_NOTIN:
@@ -3100,19 +3106,117 @@ cdbpath_motion_for_parallel_join(PlannerInfo *root,
 		case JOIN_UNIQUE_INNER:
 		case JOIN_RIGHT:
 		case JOIN_FULL:
-		case JOIN_DEDUP_SEMI:
-		case JOIN_DEDUP_SEMI_REVERSE:
 			/* Join types are not supported in parallel yet. */
 			goto fail;
+		case JOIN_DEDUP_SEMI:
+			if (!enable_parallel_dedup_semi_join)
+				goto fail;
+
+			if (!CdbPathLocus_IsPartitioned(inner.locus))
+				goto fail;
+
+			if (CdbPathLocus_IsPartitioned(outer.locus) ||
+				CdbPathLocus_IsBottleneck(outer.locus))
+			{
+				/* ok */
+			}
+			else if (CdbPathLocus_IsGeneral(outer.locus))
+			{
+				CdbPathLocus_MakeSingleQE(&outer.locus,
+										  CdbPathLocus_NumSegments(inner.locus));
+				outer.path->locus = outer.locus;
+			}
+			else if (CdbPathLocus_IsSegmentGeneral(outer.locus))
+			{
+				CdbPathLocus_MakeSingleQE(&outer.locus,
+										  CdbPathLocus_CommonSegments(inner.locus,
+																	  outer.locus));
+				outer.path->locus = outer.locus;
+			}
+			else if (CdbPathLocus_IsSegmentGeneralWorkers(outer.locus))
+			{
+				/* CBDB_PARALLEL_FIXME: Consider gather from SegmentGeneralWorkers. */
+				goto fail;
+			}
+			else
+				goto fail;
+			inner.ok_to_replicate = false;
+
+			/*
+			 * CBDB_PARALLEL:
+			 * rowidexpr is executed by 48 bits of row counter of a 64 bit int.
+			 * When in parallel mode, we need to compute the total bits of the
+			 * left 16 bits for segments and parallel workers.
+			 * The formula is:
+			 *  parallel_bits + seg_bits
+			 * while segs is max(dbid) across cluster in case that dbid segments
+			 * are uncontinuous.
+			 * And keep some room to make sure there should not be
+			 * duplicated rows when execution.
+			 */
+			if (outerParallel > 1)
+			{
+				int segs = cdbcomponent_get_maxdbid();
+				int parallel_bits = pg_leftmost_one_pos32(outerParallel) + 1;
+				int seg_bits = pg_leftmost_one_pos32(segs) + 1;
+				if (parallel_bits + seg_bits > 16)
+					goto fail;
+			}
+			outer.path = add_rowid_to_path(root, outer.path, p_rowidexpr_id);
+			*p_outer_path = outer.path;
+			break;
+
+		case JOIN_DEDUP_SEMI_REVERSE:
+			if (!enable_parallel_dedup_semi_reverse_join)
+				goto fail;
+			/* same as JOIN_DEDUP_SEMI, but with inner and outer reversed */
+			if (!CdbPathLocus_IsPartitioned(outer.locus))
+				goto fail;
+			if (CdbPathLocus_IsPartitioned(inner.locus) ||
+				CdbPathLocus_IsBottleneck(inner.locus))
+			{
+				/* ok */
+			}
+			else if (CdbPathLocus_IsGeneral(inner.locus))
+			{
+				CdbPathLocus_MakeSingleQE(&inner.locus,
+										  CdbPathLocus_NumSegments(outer.locus));
+				inner.path->locus = inner.locus;
+			}
+			else if (CdbPathLocus_IsSegmentGeneral(inner.locus))
+			{
+				CdbPathLocus_MakeSingleQE(&inner.locus,
+										  CdbPathLocus_CommonSegments(outer.locus,
+																	  inner.locus));
+				inner.path->locus = inner.locus;
+			}
+			else if (CdbPathLocus_IsSegmentGeneralWorkers(inner.locus))
+			{
+				/* CBDB_PARALLEL_FIXME: Consider gather from SegmentGeneralWorkers. */
+				goto fail;
+			}
+			else
+				goto fail;
+			outer.ok_to_replicate = false;
+			if (innerParallel > 1)
+			{
+				int segs = cdbcomponent_get_maxdbid();
+				int parallel_bits = pg_leftmost_one_pos32(innerParallel) + 1;
+				int seg_bits = pg_leftmost_one_pos32(segs) + 1;
+				if (parallel_bits + seg_bits > 16)
+					goto fail;
+			}
+			inner.path = add_rowid_to_path(root, inner.path, p_rowidexpr_id);
+			*p_inner_path = inner.path;
+			break;
+
 		default:
 			elog(ERROR, "unexpected join type %d", jointype);
 	}
 
 	/* Get rel sizes. */
 	outer.bytes = outer.path->rows * outer.path->pathtarget->width;
 	inner.bytes = inner.path->rows * inner.path->pathtarget->width;
-	int outerParallel = outer.locus.parallel_workers;
-	int innerParallel = inner.locus.parallel_workers;
 
 	if (join_quals_contain_outer_references ||
 		CdbPathLocus_IsOuterQuery(outer.locus) ||

src/backend/executor/execExpr.c

@@ -58,6 +58,7 @@
 #include "cdb/cdbvars.h"
 #include "utils/pg_locale.h"
 
+#include "port/pg_bitutils.h"
 
 typedef struct LastAttnumInfo
 {
@@ -1139,7 +1140,21 @@ ExecInitExprRec(Expr *node, ExprState *state,
 				 * value.
 				 */
 				scratch.opcode = EEOP_ROWIDEXPR;
-				scratch.d.rowidexpr.rowcounter = ((int64) GpIdentity.dbid) << 48;
+
+				/*
+				 * CBDB_PARALLEL
+				 * Planner have ensured that there is enough space for num of segments and parallel workers.
+				 * As we has not set ParallelWokerNumber yet now, use TotalParallelWorkerNumberOfSlice here
+				 * and keep bits space for ParallelWokerNumber.
+				 */
+				if (TotalParallelWorkerNumberOfSlice > 0)
+				{
+					int parallel_bits = pg_leftmost_one_pos32(TotalParallelWorkerNumberOfSlice) + 1;
+					/* Planner has checked that there is enough room. */
+					scratch.d.rowidexpr.rowcounter = ((int64) GpIdentity.dbid) << (48 + parallel_bits);
+				}
+				else
+					scratch.d.rowidexpr.rowcounter = ((int64) GpIdentity.dbid) << 48;
 
 				ExprEvalPushStep(state, &scratch);
 				break;

src/backend/executor/execExprInterp.c

@@ -1603,6 +1603,12 @@ ExecInterpExpr(ExprState *state, ExprContext *econtext, bool *isnull)
 		EEO_CASE(EEOP_ROWIDEXPR)
 		{
 			int64		rowcounter = ++op->d.rowidexpr.rowcounter;
+			/*
+			 * CBDB_PARALLEL_FIXME
+			 * Take ParallelWorkerNumberOfSlice into account for just once when initialization.
+			 */
+			if (IsParallelWorkerOfSlice())
+				rowcounter |= (((int64) ParallelWorkerNumberOfSlice) << (48));
 
 			*op->resvalue = Int64GetDatum(rowcounter);
 			*op->resnull = false;

src/backend/executor/execMain.c

@@ -1925,6 +1925,12 @@ InitPlan(QueryDesc *queryDesc, int eflags)
 			ExecSlice *sendSlice = &estate->es_sliceTable->slices[m->motionID];
 			estate->currentSliceId = sendSlice->parentIndex;
 			estate->useMppParallelMode = sendSlice->useMppParallelMode;
+			/*
+			 * CBDB_PARALLEL
+			 * Remember: parallel_workers is set to no less than = 1 when gang is filled
+			 * for convenience in Motion execution.
+			 */
+			TotalParallelWorkerNumberOfSlice = sendSlice->parallel_workers > 1 ? sendSlice->parallel_workers : 0;
 		}
 		/* Compute SubPlans' root plan nodes for SubPlans reachable from this plan root */
 		estate->locallyExecutableSubplans = getLocallyExecutableSubplans(plannedstmt, start_plan_node);
@@ -1961,9 +1967,10 @@ InitPlan(QueryDesc *queryDesc, int eflags)
 			bool		save_useMppParallelMode = estate->useMppParallelMode;
 
 			estate->currentSliceId = estate->es_plannedstmt->subplan_sliceIds[subplan_id - 1];
-			/* FIXME: test whether mpp parallel style exists for subplan case */
+			/* CBDB_PARALLEL_FIXME: test whether mpp parallel style exists for subplan case */
 			estate->useMppParallelMode = false;
 
+			/* CBDB_PARALLEL_FIXME: update TotalParallelWorkerNumberOfSlice for subplan, could it be possible? */
 			Plan	   *subplan = (Plan *) lfirst(l);
 			subplanstate = ExecInitNode(subplan, estate, sp_eflags);
 

src/backend/optimizer/path/costsize.c

@@ -6784,7 +6784,7 @@ get_parallel_divisor(Path *path)
 	 * parallel plan.
 	 */
 	/*
-	 * GPDB parallel: We don't have a leader like upstream.
+	 * CBDB_PARALLEL: We don't have a leader like upstream.
 	 * parallel_divisor is usually used to estimate rows.
 	 * Since we don't have a leader in GP parallel style, set it the same
 	 * as path's parallel_workers which may be 0 sometimes.

src/backend/optimizer/path/joinpath.c

@@ -1327,8 +1327,6 @@ sort_inner_and_outer(PlannerInfo *root,
 	if (joinrel->consider_parallel &&
 		save_jointype != JOIN_UNIQUE_OUTER &&
 		save_jointype != JOIN_FULL &&
-		save_jointype != JOIN_DEDUP_SEMI &&
-		save_jointype != JOIN_DEDUP_SEMI_REVERSE &&
 		save_jointype != JOIN_RIGHT &&
 		outerrel->partial_pathlist != NIL &&
 		bms_is_empty(joinrel->lateral_relids))
@@ -1936,8 +1934,6 @@ match_unsorted_outer(PlannerInfo *root,
 	if (joinrel->consider_parallel &&
 		save_jointype != JOIN_UNIQUE_OUTER &&
 		save_jointype != JOIN_FULL &&
-		save_jointype != JOIN_DEDUP_SEMI &&
-		save_jointype != JOIN_DEDUP_SEMI_REVERSE &&
 		save_jointype != JOIN_RIGHT &&
 		outerrel->partial_pathlist != NIL &&
 		bms_is_empty(joinrel->lateral_relids))
@@ -2032,6 +2028,9 @@ consider_parallel_nestloop(PlannerInfo *root,
 
 	if (jointype == JOIN_UNIQUE_INNER)
 		jointype = JOIN_INNER;
+
+	if (jointype == JOIN_DEDUP_SEMI || jointype == JOIN_DEDUP_SEMI_REVERSE)
+		jointype = JOIN_INNER;
 
 	foreach(lc1, outerrel->partial_pathlist)
 	{
@@ -2309,8 +2308,6 @@ hash_inner_and_outer(PlannerInfo *root,
 			save_jointype != JOIN_UNIQUE_OUTER &&
 			save_jointype != JOIN_FULL &&
 			save_jointype != JOIN_RIGHT &&
-			save_jointype != JOIN_DEDUP_SEMI &&
-			save_jointype != JOIN_DEDUP_SEMI_REVERSE &&
 			outerrel->partial_pathlist != NIL &&
 			bms_is_empty(joinrel->lateral_relids))
 		{

src/backend/utils/misc/guc_gp.c

@@ -144,6 +144,9 @@ bool		gp_appendonly_verify_write_block = false;
 bool		gp_appendonly_compaction = true;
 int			gp_appendonly_compaction_threshold = 0;
 bool		enable_parallel = false;
+bool		enable_parallel_semi_join = true;
+bool		enable_parallel_dedup_semi_join = true;
+bool		enable_parallel_dedup_semi_reverse_join = true;
 int			gp_appendonly_insert_files = 0;
 int			gp_appendonly_insert_files_tuples_range = 0;
 int			gp_random_insert_segments = 0;
@@ -3041,6 +3044,36 @@ struct config_bool ConfigureNamesBool_gp[] =
 		false,
 		NULL, NULL, NULL
 	},
+	{
+		{"enable_parallel_semi_join", PGC_USERSET, DEVELOPER_OPTIONS,
+			gettext_noop("allow to use of parallel semi join."),
+			NULL,
+			GUC_EXPLAIN
+		},
+		&enable_parallel_semi_join,
+		true,
+		NULL, NULL, NULL
+	},
+	{
+		{"enable_parallel_dedup_semi_join", PGC_USERSET, DEVELOPER_OPTIONS,
+			gettext_noop("allow to use of parallel dedup semi join."),
+			NULL,
+			GUC_EXPLAIN
+		},
+		&enable_parallel_dedup_semi_join,
+		true,
+		NULL, NULL, NULL
+	},
+	{
+		{"enable_parallel_dedup_semi_reverse_join", PGC_USERSET, DEVELOPER_OPTIONS,
+			gettext_noop("allow to use of parallel dedup semi reverse join."),
+			NULL,
+			GUC_EXPLAIN
+		},
+		&enable_parallel_dedup_semi_reverse_join,
+		true,
+		NULL, NULL, NULL
+	},
 	{
 		{"gp_internal_is_singlenode", PGC_POSTMASTER, UNGROUPED,
 			 gettext_noop("Is in SingleNode mode (no segments). WARNING: user SHOULD NOT set this by any means."),

src/include/access/parallel.h

@@ -63,6 +63,10 @@ extern volatile bool ParallelMessagePending;
 extern PGDLLIMPORT int ParallelWorkerNumber;
 extern PGDLLIMPORT bool InitializingParallelWorker;
 
+/* CBDB_PARALLEL: Total parallel workers of a slice including myself, 0 for no parallel */
+extern PGDLLIMPORT int ParallelWorkerNumberOfSlice;
+extern PGDLLIMPORT int TotalParallelWorkerNumberOfSlice;
+
 typedef struct ParallelEntryTag
 {
 	int cid;
@@ -90,6 +94,15 @@ typedef struct GpParallelDSMEntry
 	Barrier		build_barrier;	/* synchronization for the build dsm phases */
 } GpParallelDSMEntry;
 
+/*
+ * CBDB_PARALLEL
+ * The Postgres uses ParallelWorkerNumber to handle background workers including
+ * parallel workers under Gather node.
+ * To avoid mixing them and assertion failure, we use ParallelWorkerNumberOfSlice
+ * to indentify CBDB style parallel mode.
+ */
+#define		IsParallelWorkerOfSlice()	(ParallelWorkerNumberOfSlice >= 0)
+
 #define		IsParallelWorker()		(ParallelWorkerNumber >= 0)
 
 extern ParallelContext *CreateParallelContext(const char *library_name,

src/include/utils/guc.h

@@ -290,6 +290,9 @@ extern bool gp_appendonly_verify_block_checksums;
 extern bool gp_appendonly_verify_write_block;
 extern bool gp_appendonly_compaction;
 extern bool enable_parallel;
+extern bool enable_parallel_semi_join;
+extern bool enable_parallel_dedup_semi_join;
+extern bool enable_parallel_dedup_semi_reverse_join;
 extern int  gp_appendonly_insert_files;
 extern int  gp_appendonly_insert_files_tuples_range;
 extern int  gp_random_insert_segments;

src/include/utils/unsync_guc_name.h

@@ -125,6 +125,9 @@
 		"enable_nestloop",
 		"enable_parallel_append",
 		"enable_parallel_hash",
+		"enable_parallel_semi_join",
+		"enable_parallel_dedup_semi_join",
+		"enable_parallel_dedup_semi_reverse_join",
 		"enable_partition_pruning",
 		"enable_partitionwise_aggregate",
 		"enable_partitionwise_join",