greenplumn cdbsetop 源码

  • 2022-08-18
  • 浏览 (325)

greenplumn cdbsetop 代码

文件路径:/src/backend/cdb/cdbsetop.c

/*-------------------------------------------------------------------------
 *
 * cdbsetop.c
 *	  Routines to aid in planning set-operation queries for parallel
 *    execution.  This is, essentially, an extension of the file
 *    optimizer/prep/prepunion.c, although some functions are not
 *    externalized.
 *
 * Portions Copyright (c) 2005-2008, Greenplum inc
 * Portions Copyright (c) 2012-Present VMware, Inc. or its affiliates.
 * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
 * Portions Copyright (c) 1994, Regents of the University of California
 *
 *
 * IDENTIFICATION
 *	    src/backend/cdb/cdbsetop.c
 *
 *-------------------------------------------------------------------------
 */
#include "postgres.h"

#include "nodes/makefuncs.h"
#include "optimizer/planmain.h"

#include "cdb/cdbhash.h"
#include "cdb/cdbmutate.h"
#include "cdb/cdbpath.h"
#include "cdb/cdbsetop.h"
#include "cdb/cdbutil.h"
#include "cdb/cdbvars.h"


/*
 * Function: choose_setop_type
 *
 * Decide what type of plan to use for a set operation based on the loci of
 * the node list input to the set operation.
 *
 * See the comments in cdbsetop.h for discussion of types of setop plan.
 */
GpSetOpType
choose_setop_type(List *pathlist)
{
	ListCell   *cell;
	bool		ok_general = true;
	bool		ok_partitioned = true;
	bool		ok_single_qe = true;
	bool		has_partitioned = false;

	Assert(Gp_role == GP_ROLE_DISPATCH || Gp_role == GP_ROLE_UTILITY);

	foreach(cell, pathlist)
	{
		Path	   *subpath = (Path *) lfirst(cell);

		switch (subpath->locus.locustype)
		{
			case CdbLocusType_Hashed:
			case CdbLocusType_HashedOJ:
			case CdbLocusType_Strewn:
				ok_general = false;
				has_partitioned = true;
				break;

			case CdbLocusType_Entry:
				ok_general = ok_partitioned = ok_single_qe = false;
				break;

			case CdbLocusType_OuterQuery:
				return PSETOP_SEQUENTIAL_OUTERQUERY;

			case CdbLocusType_SingleQE:
				ok_general = false;
				break;

			case CdbLocusType_SegmentGeneral:
				ok_general = false;
				break;

			case CdbLocusType_General:
				break;

			case CdbLocusType_Replicated:
				break;

			case CdbLocusType_Null:
				elog(ERROR, "unexpected Null locus in set operation branch");
				break;
			default:
				elog(ERROR, "unexpected locus type in set operation branch");
				break;
		}
	}

	if (ok_general)
		return PSETOP_GENERAL;
	else if (ok_partitioned && has_partitioned)
		return PSETOP_PARALLEL_PARTITIONED;
	else if (ok_single_qe)
		return PSETOP_SEQUENTIAL_QE;

	return PSETOP_SEQUENTIAL_QD;
}


void
adjust_setop_arguments(PlannerInfo *root, List *pathlist, List *tlist_list, GpSetOpType setop_type)
{
	ListCell   *pathcell;
	ListCell   *tlistcell;
	Path	   *adjusted_path;
	CdbPathLocus locus;

	forboth(pathcell, pathlist, tlistcell, tlist_list)
	{
		Path	   *subpath = (Path *) lfirst(pathcell);
		List	   *subtlist = (List *) lfirst(tlistcell);

		adjusted_path = subpath;
		switch (setop_type)
		{
			case PSETOP_GENERAL:
				/* This only occurs when all arguments are general. */
				break;

			case PSETOP_PARALLEL_PARTITIONED:
				switch (subpath->locus.locustype)
				{
					case CdbLocusType_Hashed:
					case CdbLocusType_HashedOJ:
					case CdbLocusType_Strewn:
						break;
					case CdbLocusType_SingleQE:
					case CdbLocusType_General:
					case CdbLocusType_SegmentGeneral:
						/*
						 * The setop itself will run on an N-gang, so we need
						 * to arrange for the singleton input to be separately
						 * dispatched to a 1-gang and collect its result on
						 * one of our N QEs. Hence ...
						 */
						adjusted_path = make_motion_hash_all_targets(root, subpath, subtlist);
						break;
					case CdbLocusType_Null:
					case CdbLocusType_Entry:
					case CdbLocusType_Replicated:
					case CdbLocusType_OuterQuery:
					case CdbLocusType_End:
						ereport(ERROR, (errcode(ERRCODE_INTERNAL_ERROR),
										errmsg("unexpected argument locus to set operation")));
						break;
				}
				break;

			case PSETOP_SEQUENTIAL_QD:
				switch (subpath->locus.locustype)
				{
					case CdbLocusType_Hashed:
					case CdbLocusType_HashedOJ:
					case CdbLocusType_Strewn:
						CdbPathLocus_MakeEntry(&locus);
						adjusted_path = cdbpath_create_motion_path(root, subpath, NULL, false,
																   locus);
						break;

					case CdbLocusType_SingleQE:
					case CdbLocusType_SegmentGeneral:
						/*
						 * The input was focused on a single QE, but we need it in the QD.
						 * It's bit silly to add a Motion to just move the whole result from
						 * single QE to QD, it would be better to produce the result in the
						 * QD in the first place, and avoid the Motion. But it's too late
						 * to modify the subpath.
						 */
						CdbPathLocus_MakeEntry(&locus);
						adjusted_path = cdbpath_create_motion_path(root, subpath, NULL, false,
																   locus);
						break;

					case CdbLocusType_Entry:
					case CdbLocusType_General:
					case CdbLocusType_OuterQuery:
						break;

					case CdbLocusType_Null:
					case CdbLocusType_Replicated:
					case CdbLocusType_End:
						ereport(ERROR, (errcode(ERRCODE_INTERNAL_ERROR),
										errmsg("unexpected argument locus to set operation")));
						break;
				}
				break;

			case PSETOP_SEQUENTIAL_QE:
			case PSETOP_SEQUENTIAL_OUTERQUERY:
				switch (subpath->locus.locustype)
				{
					case CdbLocusType_Hashed:
					case CdbLocusType_HashedOJ:
					case CdbLocusType_Strewn:
						/* Gather to QE.  No need to keep ordering. */
						CdbPathLocus_MakeSingleQE(&locus, getgpsegmentCount());
						adjusted_path = cdbpath_create_motion_path(root, subpath, NULL, false,
																   locus);
						break;

					case CdbLocusType_SingleQE:
						break;

					case CdbLocusType_OuterQuery:
					case CdbLocusType_General:
						break;

					case CdbLocusType_SegmentGeneral:
						/* Gather to QE.  No need to keep ordering. */
						CdbPathLocus_MakeSingleQE(&locus, getgpsegmentCount());
						adjusted_path = cdbpath_create_motion_path(root, subpath, NULL, false,
																   locus);
						break;

					case CdbLocusType_Entry:
					case CdbLocusType_Null:
					case CdbLocusType_Replicated:
					case CdbLocusType_End:
						ereport(ERROR, (errcode(ERRCODE_INTERNAL_ERROR),
										errmsg("unexpected argument locus to set operation")));
						break;
				}
				break;

			default:
				/* Can't happen! */
				ereport(ERROR, (errcode(ERRCODE_INTERNAL_ERROR),
								errmsg("unexpected arguments to set operation")));
				break;
		}

		/* If we made changes, inject them into the argument list. */
		if (subpath != adjusted_path)
		{
			subpath = adjusted_path;
			pathcell->data.ptr_value = subpath;
		}
	}

	return;
}

/*
 * make_motion_hash_all_targets
 *		Add a Motion node atop the given subplath to hash collocate
 *      tuples non-distinct on the non-junk attributes.  This motion
 *      should only be applied to a non-replicated, non-root subpath.
 *
 * This will align with the sort attributes used as input to a SetOp
 * or Unique operator. This is used in path for UNION and other
 * set-operations that implicitly do a DISTINCT on the whole target
 * list.
 */
Path *
make_motion_hash_all_targets(PlannerInfo *root, Path *subpath, List *tlist)
{
	ListCell   *cell;
	List	   *hashexprs = NIL;
	List	   *hashopfamilies = NIL;
	List	   *hashsortrefs = NIL;
	CdbPathLocus locus;

	foreach(cell, tlist)
	{
		TargetEntry *tle = (TargetEntry *) lfirst(cell);
		Oid			opfamily;

		if (tle->resjunk)
			continue;

		opfamily = cdb_default_distribution_opfamily_for_type(exprType((Node *) tle->expr));
		if (!opfamily)
			continue;		/* not hashable */

		hashexprs = lappend(hashexprs, copyObject(tle->expr));
		hashopfamilies = lappend_oid(hashopfamilies, opfamily);
		hashsortrefs = lappend_int(hashsortrefs, tle->ressortgroupref);
	}

	if (hashexprs)
	{
		/* Distribute to ALL to maximize parallelism */
		locus = cdbpathlocus_from_exprs(root,
										subpath->parent,
										hashexprs,
										hashopfamilies,
										hashsortrefs,
										getgpsegmentCount());
	}
	else
	{
		/*
		 * Degenerate case, where none of the columns are hashable.
		 *
		 * (If the caller knew this, it probably would have been better to
		 * produce a different plan, with Sorts in the segments, and an
		 * order-preserving gather on the top.)
		 */
		CdbPathLocus_MakeSingleQE(&locus, getgpsegmentCount());
	}

	return cdbpath_create_motion_path(root, subpath, subpath->pathkeys,
									  false, locus);
}

/*
 *     Marks an Append plan with its locus based on the set operation
 *     type determined during examination of the arguments.
 */
void
mark_append_locus(Path *path, GpSetOpType optype)
{
	switch (optype)
	{
		case PSETOP_GENERAL:
			CdbPathLocus_MakeGeneral(&path->locus);
			break;
		case PSETOP_PARALLEL_PARTITIONED:
			CdbPathLocus_MakeStrewn(&path->locus, getgpsegmentCount());
			break;
		case PSETOP_SEQUENTIAL_QD:
			CdbPathLocus_MakeEntry(&path->locus);
			break;
		case PSETOP_SEQUENTIAL_QE:
			CdbPathLocus_MakeSingleQE(&path->locus, getgpsegmentCount());
			break;
		case PSETOP_SEQUENTIAL_OUTERQUERY:
			CdbPathLocus_MakeOuterQuery(&path->locus);
			break;
		case PSETOP_NONE:
			break;
	}
}

相关信息

greenplumn 源码目录

相关文章

greenplumn cdbappendonlystorageformat 源码

greenplumn cdbappendonlystorageread 源码

greenplumn cdbappendonlystoragewrite 源码

greenplumn cdbappendonlyxlog 源码

greenplumn cdbbufferedappend 源码

greenplumn cdbbufferedread 源码

greenplumn cdbcat 源码

greenplumn cdbcopy 源码

greenplumn cdbdistributedsnapshot 源码

greenplumn cdbdistributedxacts 源码

0  赞