greenplumn cdbpath 源码
greenplumn cdbpath 代码
文件路径：/src/backend/cdb/cdbpath.c
/*-------------------------------------------------------------------------
*
* cdbpath.c
*
* Portions Copyright (c) 2005-2008, Greenplum inc
* Portions Copyright (c) 2012-Present VMware, Inc. or its affiliates.
*
*
* IDENTIFICATION
* src/backend/cdb/cdbpath.c
*
*-------------------------------------------------------------------------
*/
#include "postgres.h"
#include "access/hash.h"
#include "access/heapam.h"
#include "access/htup_details.h"
#include "catalog/pg_amop.h"
#include "catalog/pg_opclass.h"
#include "catalog/pg_operator.h"
#include "catalog/pg_trigger.h"
#include "commands/trigger.h"
#include "nodes/makefuncs.h" /* makeFuncExpr() */
#include "nodes/nodeFuncs.h" /* exprType() */
#include "nodes/pathnodes.h" /* PlannerInfo, RelOptInfo */
#include "optimizer/cost.h" /* set_rel_width() */
#include "optimizer/optimizer.h" /* cpu_tuple_cost */
#include "optimizer/pathnode.h" /* Path, pathnode_walker() */
#include "optimizer/paths.h"
#include "optimizer/planmain.h"
#include "optimizer/tlist.h"
#include "parser/parsetree.h"
#include "parser/parse_expr.h" /* exprType() */
#include "parser/parse_oper.h"
#include "utils/catcache.h"
#include "utils/guc.h"
#include "utils/lsyscache.h"
#include "utils/syscache.h"
#include "cdb/cdbdef.h" /* CdbSwap() */
#include "cdb/cdbhash.h"
#include "cdb/cdbpath.h" /* me */
#include "cdb/cdbutil.h"
#include "cdb/cdbvars.h"
typedef struct
{
CdbPathLocus locus;
CdbPathLocus move_to;
double bytes;
Path *path;
List *pathkeys;
bool ok_to_replicate;
bool require_existing_order;
bool has_wts; /* Does the rel have WorkTableScan? */
} CdbpathMfjRel;
static bool try_redistribute(PlannerInfo *root, CdbpathMfjRel *g,
CdbpathMfjRel *o, List *redistribution_clauses);
static SplitUpdatePath *make_splitupdate_path(PlannerInfo *root, Path *subpath, Index rti);
static bool can_elide_explicit_motion(PlannerInfo *root, Index rti, Path *subpath, GpPolicy *policy);
/*
* cdbpath_cost_motion
* Fills in the cost estimate fields in a MotionPath node.
*/
void
cdbpath_cost_motion(PlannerInfo *root, CdbMotionPath *motionpath)
{
Path *subpath = motionpath->subpath;
Cost cost_per_row;
Cost motioncost;
double recvrows;
double sendrows;
double send_segments;
double recv_segments;
double total_rows;
if (CdbPathLocus_IsPartitioned(motionpath->path.locus))
recv_segments = CdbPathLocus_NumSegments(motionpath->path.locus);
else
recv_segments = 1;
if (CdbPathLocus_IsPartitioned(subpath->locus))
send_segments = CdbPathLocus_NumSegments(subpath->locus);
else
send_segments = 1;
/*
* Estimate the total number of rows being sent.
*
* The base estimate is computed by multiplying the subpath's rows with
* the number of sending segments. But in some cases, that leads to too
* large estimates, if the subpath's estimate was "clamped" to 1 row. The
* typical example is a single-row select like "SELECT * FROM table WHERE
* key = 123. The Scan on the table returns only one row, on one segment,
* and the estimate on the Scan node is 1 row. If you have e.g. 3
* segments, and we just multiplied the subpath's row estimate by 3, we
* would estimate that the Gather returns 3 rows, even though there is
* only one matching row in the table. Using the 'rows' estimate on the
* RelOptInfo is more accurate in such cases. To correct that, if the
* subpath's estimate is 1 row, but the underlying relation's estimate is
* smaller, use the underlying relation's estimate.
*
* We don't always use the relation's estimate, because there might be
* nodes like ProjectSet or Limit in the subpath, in which case the
* subpath's estimate is more accurate. Also, the relation might not have
* a valid 'rows' estimate; upper rels, for example, do not. So check for
* that too.
*/
total_rows = subpath->rows * send_segments;
if (subpath->rows == 1.0 &&
motionpath->path.parent->rows > 0 &&
motionpath->path.parent->rows < total_rows)
{
/* use the RelOptInfo's estimate */
total_rows = motionpath->path.parent->rows;
}
motionpath->path.rows = clamp_row_est(total_rows / recv_segments);
cost_per_row = (gp_motion_cost_per_row > 0.0)
? gp_motion_cost_per_row
: 2.0 * cpu_tuple_cost;
sendrows = subpath->rows;
recvrows = motionpath->path.rows;
motioncost = cost_per_row * 0.5 * (sendrows + recvrows);
motionpath->path.total_cost = motioncost + subpath->total_cost;
motionpath->path.startup_cost = subpath->startup_cost;
motionpath->path.memory = subpath->memory;
} /* cdbpath_cost_motion */
/*
* cdbpath_create_motion_path
* Returns a Path that delivers the subpath result to the
* given locus, or NULL if it can't be done.
*
* 'pathkeys' must specify an ordering equal to or weaker than the
* subpath's existing ordering.
*
* If no motion is needed, the caller's subpath is returned unchanged.
* Else if require_existing_order is true, NULL is returned if the
* motion would not preserve an ordering at least as strong as the
* specified ordering; also NULL is returned if pathkeys is NIL
* meaning the caller is just checking and doesn't want to add motion.
* Else a CdbMotionPath is returned having either the specified pathkeys
* (if given and the motion uses Merge Receive), or the pathkeys
* of the original subpath (if the motion is order-preserving), or no
* pathkeys otherwise (the usual case).
*/
Path *
cdbpath_create_motion_path(PlannerInfo *root,
Path *subpath,
List *pathkeys,
bool require_existing_order,
CdbPathLocus locus)
{
CdbMotionPath *pathnode;
Assert(cdbpathlocus_is_valid(locus) &&
cdbpathlocus_is_valid(subpath->locus));
/*
* Motion is to change path's locus, if target locus is the
* same as the subpath's, there is no need to add motion.
*/
if (cdbpathlocus_equal(subpath->locus, locus))
return subpath;
/* Moving subpath output to a single executor process (qDisp or qExec)? */
if (CdbPathLocus_IsOuterQuery(locus))
{
/* Outer -> Outer is a no-op */
if (CdbPathLocus_IsOuterQuery(subpath->locus))
{
return subpath;
}
if (CdbPathLocus_IsGeneral(subpath->locus))
{
/* XXX: this is a bit bogus. We just change the subpath's locus. */
subpath->locus = locus;
return subpath;
}
if (CdbPathLocus_IsEntry(subpath->locus) ||
CdbPathLocus_IsSingleQE(subpath->locus))
{
/*
* XXX: this is a bit bogus. We just change the subpath's locus.
*
* This is also bogus, because the outer query might need to run
* in segments.
*/
subpath->locus = locus;
return subpath;
}
}
else if (CdbPathLocus_IsBottleneck(locus))
{
/* entry-->entry? No motion needed. */
if (CdbPathLocus_IsEntry(subpath->locus) &&
CdbPathLocus_IsEntry(locus))
{
return subpath;
}
/* singleQE-->singleQE? No motion needed. */
if (CdbPathLocus_IsSingleQE(subpath->locus) &&
CdbPathLocus_IsSingleQE(locus))
{
subpath->locus.numsegments = CdbPathLocus_CommonSegments(subpath->locus, locus);
return subpath;
}
/* entry-->singleQE? Don't move. Slice's QE will run on entry db. */
if (CdbPathLocus_IsEntry(subpath->locus))
{
return subpath;
}
/* outerquery-->entry? No motion needed. */
if (CdbPathLocus_IsOuterQuery(subpath->locus) &&
CdbPathLocus_IsEntry(locus))
{
return subpath;
}
/* singleQE-->entry? Don't move. Slice's QE will run on entry db. */
if (CdbPathLocus_IsSingleQE(subpath->locus))
{
/*
* If the subpath requires parameters, we cannot generate Motion atop of it.
*/
if (!bms_is_empty(PATH_REQ_OUTER(subpath)))
return NULL;
/*
* Create CdbMotionPath node to indicate that the slice must be
* dispatched to a singleton gang running on the entry db. We
* merely use this node to note that the path has 'Entry' locus;
* no corresponding Motion node will be created in the Plan tree.
*/
Assert(CdbPathLocus_IsEntry(locus));
pathnode = makeNode(CdbMotionPath);
pathnode->path.pathtype = T_Motion;
pathnode->path.parent = subpath->parent;
/* Motion doesn't project, so use source path's pathtarget */
pathnode->path.pathtarget = subpath->pathtarget;
pathnode->path.locus = locus;
pathnode->path.rows = subpath->rows;
/* GPDB_96_MERGE_FIXME: When is a Motion path parallel-safe? I tried
* setting this to 'false' initially, to play it safe, but somehow
* the Paths with motions ended up in gather plans anyway, and tripped
* assertion failures.
*/
pathnode->path.parallel_aware = false;
pathnode->path.parallel_safe = subpath->parallel_safe;
pathnode->path.parallel_workers = subpath->parallel_workers;
pathnode->path.pathkeys = pathkeys;
pathnode->subpath = subpath;
/* Costs, etc, are same as subpath. */
pathnode->path.startup_cost = subpath->total_cost;
pathnode->path.total_cost = subpath->total_cost;
pathnode->path.memory = subpath->memory;
pathnode->path.motionHazard = subpath->motionHazard;
/* Motion nodes are never rescannable. */
pathnode->path.rescannable = false;
return (Path *) pathnode;
}
if (CdbPathLocus_IsSegmentGeneral(subpath->locus) ||
CdbPathLocus_IsReplicated(subpath->locus))
{
/*
* If the subpath requires parameters, we cannot generate Motion atop of it.
*/
if (!bms_is_empty(PATH_REQ_OUTER(subpath)))
return NULL;
/*
* Data is only available on segments, to distingush it with
* CdbLocusType_General, adding a motion to indicated this
* slice must be executed on a singleton gang.
*
* This motion may be redundant for segmentGeneral --> singleQE
* if the singleQE is not promoted to executed on qDisp in the
* end, so in cdbllize_fix_outer_query_motions(), we will omit it.
*/
pathnode = makeNode(CdbMotionPath);
pathnode->path.pathtype = T_Motion;
pathnode->path.parent = subpath->parent;
/* Motion doesn't project, so use source path's pathtarget */
pathnode->path.pathtarget = subpath->pathtarget;
pathnode->path.locus = locus;
pathnode->path.rows = subpath->rows;
pathnode->path.pathkeys = pathkeys;
/* GPDB_96_MERGE_FIXME: When is a Motion path parallel-safe? I tried
* setting this to 'false' initially, to play it safe, but somehow
* the Paths with motions ended up in gather plans anyway, and tripped
* assertion failures.
*/
pathnode->path.parallel_aware = false;
pathnode->path.parallel_safe = subpath->parallel_safe;
pathnode->path.parallel_workers = subpath->parallel_workers;
pathnode->subpath = subpath;
/* Costs, etc, are same as subpath. */
pathnode->path.startup_cost = subpath->total_cost;
pathnode->path.total_cost = subpath->total_cost;
pathnode->path.memory = subpath->memory;
pathnode->path.motionHazard = subpath->motionHazard;
/* Motion nodes are never rescannable. */
pathnode->path.rescannable = false;
return (Path *) pathnode;
}
/* No motion needed if subpath can run anywhere giving same output. */
if (CdbPathLocus_IsGeneral(subpath->locus))
{
/*
* general-->(entry|singleqe), no motion is needed, can run
* directly on any of the common segments
*/
return subpath;
}
/* Fail if caller refuses motion. */
if (require_existing_order &&
!pathkeys)
return NULL;
/*
* Must be partitioned-->singleton. If caller gave pathkeys, they'll
* be used for Merge Receive. If no pathkeys, Union Receive will
* arbitrarily interleave the rows from the subpath partitions in no
* special order.
*/
if (!CdbPathLocus_IsPartitioned(subpath->locus))
goto invalid_motion_request;
}
/* Output from a single process to be distributed over a gang? */
else if (CdbPathLocus_IsBottleneck(subpath->locus))
{
/* Must be bottleneck-->partitioned or bottleneck-->replicated */
if (!CdbPathLocus_IsPartitioned(locus) &&
!CdbPathLocus_IsReplicated(locus))
goto invalid_motion_request;
/* Fail if caller disallows motion. */
if (require_existing_order &&
!pathkeys)
return NULL;
/* Each qExec receives a subset of the rows, with ordering preserved. */
pathkeys = subpath->pathkeys;
}
/* Redistributing partitioned subpath output from one gang to another? */
else if (CdbPathLocus_IsPartitioned(subpath->locus))
{
/* partitioned-->partitioned? */
if (CdbPathLocus_IsPartitioned(locus))
{
/* No motion if subpath partitioning matches caller's request. */
if (cdbpathlocus_equal(subpath->locus, locus))
return subpath;
}
/* Must be partitioned-->replicated */
else if (!CdbPathLocus_IsReplicated(locus))
goto invalid_motion_request;
/* Fail if caller insists on ordered result or no motion. */
if (require_existing_order)
return NULL;
/*
* Output streams lose any ordering they had. Only a qDisp or
* singleton qExec can merge sorted streams (for now).
*/
pathkeys = NIL;
}
/* If subplan uses no tables, it can run on qDisp or a singleton qExec. */
else if (CdbPathLocus_IsGeneral(subpath->locus))
{
/*
* No motion needed if general-->general or general-->replicated or
* general-->segmentGeneral
*/
if (CdbPathLocus_IsGeneral(locus) ||
CdbPathLocus_IsReplicated(locus) ||
CdbPathLocus_IsSegmentGeneral(locus))
{
return subpath;
}
/* Must be general-->partitioned. */
if (!CdbPathLocus_IsPartitioned(locus))
goto invalid_motion_request;
/* Fail if caller wants no motion. */
if (require_existing_order &&
!pathkeys)
return NULL;
/* Since the motion is 1-to-many, the rows remain in the same order. */
pathkeys = subpath->pathkeys;
}
/* Does subpath produce same multiset of rows on every qExec of its gang? */
else if (CdbPathLocus_IsReplicated(subpath->locus))
{
/* No-op if replicated-->replicated. */
if (CdbPathLocus_IsReplicated(locus))
{
Assert(CdbPathLocus_NumSegments(locus) <=
CdbPathLocus_NumSegments(subpath->locus));
subpath->locus.numsegments = CdbPathLocus_NumSegments(locus);
return subpath;
}
/* Other destinations aren't used or supported at present. */
goto invalid_motion_request;
}
/* Most motions from SegmentGeneral (replicated table) are disallowed */
else if (CdbPathLocus_IsSegmentGeneral(subpath->locus))
{
/*
* The only allowed case is a SegmentGeneral to Hashed motion,
* and SegmentGeneral's numsegments is smaller than Hashed's.
* In such a case we redistribute SegmentGeneral to Hashed.
*
* FIXME: HashedOJ?
*/
if (CdbPathLocus_IsPartitioned(locus))
{
pathkeys = subpath->pathkeys;
}
else if (CdbPathLocus_IsReplicated(locus))
{
if (CdbPathLocus_NumSegments(locus) <= CdbPathLocus_NumSegments(subpath->locus))
{
subpath->locus.numsegments = CdbPathLocus_NumSegments(locus);
return subpath;
}
else
{
pathkeys = subpath->pathkeys;
}
}
else if (CdbPathLocus_IsSegmentGeneral(locus))
{
subpath->locus.numsegments = Min(subpath->locus.numsegments, locus.numsegments);
return subpath;
}
else
goto invalid_motion_request;
}
else
goto invalid_motion_request;
/* Don't materialize before motion. */
if (IsA(subpath, MaterialPath))
subpath = ((MaterialPath *) subpath)->subpath;
/*
* MPP-3300: materialize *before* motion can never help us, motion pushes
* data. other nodes pull. We relieve motion deadlocks by adding
* materialize nodes on top of motion nodes
*/
if(IsA(subpath,SubqueryScanPath)
&& CdbPathLocus_IsBottleneck(locus)
&& !subpath->pathkeys
&& ((SubqueryScanPath *)subpath)->subpath->pathkeys)
{
/*
* In gpdb, when there is a Gather Motion on top of a SubqueryScan,
* it is hard to keep the sort order information. The subquery's
* path key cannot be pulled up, if the parent query doesn't have
* an equivalence class corresponding to the subquery's sort order.
*
* e.g. create a view with an ORDER BY:
*
* CREATE VIEW v AS SELECT va, vn FROM sourcetable ORDER BY vn;
*
* and query it:
*
* SELECT va FROM v_sourcetable;
*
* So, push down the Gather Motion if the SubqueryScan dose not
* have pathkey but the SubqueryScan's subpath does.
*
*/
SubqueryScanPath *subqueryScanPath = (SubqueryScanPath *)subpath;
SubqueryScanPath *newSubqueryScanPath = NULL;
Path *motionPath = NULL;
subpath = subqueryScanPath->subpath;
motionPath = cdbpath_create_motion_path(root,
subpath,
subpath->pathkeys,
true,
locus);
newSubqueryScanPath = create_subqueryscan_path(root,
subqueryScanPath->path.parent,
motionPath,
subqueryScanPath->path.pathkeys,
locus,
subqueryScanPath->required_outer);
/* apply final path target */
newSubqueryScanPath = (SubqueryScanPath *)apply_projection_to_path(root,
subqueryScanPath->path.parent,
(Path *) newSubqueryScanPath,
subqueryScanPath->path.pathtarget);
return (Path *) newSubqueryScanPath;
}
/*
* If the subpath requires parameters, we cannot generate Motion atop of it.
*/
if (!bms_is_empty(PATH_REQ_OUTER(subpath)))
return NULL;
/* Create CdbMotionPath node. */
pathnode = makeNode(CdbMotionPath);
pathnode->path.pathtype = T_Motion;
pathnode->path.parent = subpath->parent;
/* Motion doesn't project, so use source path's pathtarget */
pathnode->path.pathtarget = subpath->pathtarget;
pathnode->path.locus = locus;
pathnode->path.rows = subpath->rows;
pathnode->path.pathkeys = pathkeys;
/* GPDB_96_MERGE_FIXME: When is a Motion path parallel-safe? I tried
* setting this to 'false' initially, to play it safe, but somehow
* the Paths with motions ended up in gather plans anyway, and tripped
* assertion failures.
*/
pathnode->path.parallel_aware = false;
pathnode->path.parallel_safe = subpath->parallel_safe;
pathnode->path.parallel_workers = subpath->parallel_workers;
pathnode->subpath = subpath;
pathnode->is_explicit_motion = false;
/* Cost of motion */
cdbpath_cost_motion(root, pathnode);
/* Tell operators above us that slack may be needed for deadlock safety. */
pathnode->path.motionHazard = true;
pathnode->path.rescannable = false;
/*
* A motion to bring data to the outer query's locus needs a Material node
* on top, to shield the Motion node from rescanning, when the SubPlan
* is rescanned.
*/
if (CdbPathLocus_IsOuterQuery(locus))
{
return (Path *) create_material_path(root, subpath->parent,
&pathnode->path);
}
return (Path *) pathnode;
/* Unexpected source or destination locus. */
invalid_motion_request:
elog(ERROR, "could not build Motion path");
return NULL;
} /* cdbpath_create_motion_path */
Path *
cdbpath_create_explicit_motion_path(PlannerInfo *root,
Path *subpath,
CdbPathLocus locus)
{
CdbMotionPath *pathnode;
/* Create CdbMotionPath node. */
pathnode = makeNode(CdbMotionPath);
pathnode->path.pathtype = T_Motion;
pathnode->path.parent = subpath->parent;
/* Motion doesn't project, so use source path's pathtarget */
pathnode->path.pathtarget = subpath->pathtarget;
pathnode->path.locus = locus;
pathnode->path.rows = subpath->rows;
pathnode->path.pathkeys = NIL;
/* GPDB_96_MERGE_FIXME: When is a Motion path parallel-safe? I tried
* setting this to 'false' initially, to play it safe, but somehow
* the Paths with motions ended up in gather plans anyway, and tripped
* assertion failures.
*/
pathnode->path.parallel_aware = false;
pathnode->path.parallel_safe = subpath->parallel_safe;
pathnode->path.parallel_workers = subpath->parallel_workers;
pathnode->subpath = subpath;
pathnode->is_explicit_motion = true;
/* Cost of motion */
cdbpath_cost_motion(root, pathnode);
/* Tell operators above us that slack may be needed for deadlock safety. */
pathnode->path.motionHazard = true;
pathnode->path.rescannable = false;
return (Path *) pathnode;
}
Path *
cdbpath_create_broadcast_motion_path(PlannerInfo *root,
Path *subpath,
int numsegments)
{
CdbMotionPath *pathnode;
/* Create CdbMotionPath node. */
pathnode = makeNode(CdbMotionPath);
pathnode->path.pathtype = T_Motion;
pathnode->path.parent = subpath->parent;
/* Motion doesn't project, so use source path's pathtarget */
pathnode->path.pathtarget = subpath->pathtarget;
CdbPathLocus_MakeReplicated(&pathnode->path.locus, numsegments);
pathnode->path.rows = subpath->rows;
pathnode->path.pathkeys = NIL;
/* GPDB_96_MERGE_FIXME: When is a Motion path parallel-safe? I tried
* setting this to 'false' initially, to play it safe, but somehow
* the Paths with motions ended up in gather plans anyway, and tripped
* assertion failures.
*/
pathnode->path.parallel_aware = false;
pathnode->path.parallel_safe = subpath->parallel_safe;
pathnode->path.parallel_workers = subpath->parallel_workers;
pathnode->subpath = subpath;
pathnode->is_explicit_motion = false;
/* Cost of motion */
cdbpath_cost_motion(root, pathnode);
/* Tell operators above us that slack may be needed for deadlock safety. */
pathnode->path.motionHazard = true;
pathnode->path.rescannable = false;
return (Path *) pathnode;
}
/*
*/
static CdbMotionPath *
make_motion_path(PlannerInfo *root, Path *subpath,
CdbPathLocus locus,
bool is_explicit_motion,
GpPolicy *policy)
{
CdbMotionPath *pathnode;
/* Create CdbMotionPath node. */
pathnode = makeNode(CdbMotionPath);
pathnode->path.pathtype = T_Motion;
pathnode->path.parent = subpath->parent;
/* Motion doesn't project, so use source path's pathtarget */
pathnode->path.pathtarget = subpath->pathtarget;
pathnode->path.locus = locus;
pathnode->path.rows = subpath->rows;
pathnode->path.pathkeys = NIL;
/* GPDB_96_MERGE_FIXME: When is a Motion path parallel-safe? I tried
* setting this to 'false' initially, to play it safe, but somehow
* the Paths with motions ended up in gather plans anyway, and tripped
* assertion failures.
*/
pathnode->path.parallel_aware = false;
pathnode->path.parallel_safe = subpath->parallel_safe;
pathnode->path.parallel_workers = subpath->parallel_workers;
pathnode->subpath = subpath;
pathnode->is_explicit_motion = is_explicit_motion;
pathnode->policy = policy;
/* Cost of motion */
cdbpath_cost_motion(root, pathnode);
/* Tell operators above us that slack may be needed for deadlock safety. */
pathnode->path.motionHazard = true;
pathnode->path.rescannable = false;
return pathnode;
}
/*
* cdbpath_match_preds_to_partkey_tail
*
* Returns true if all of the locus's partitioning key expressions are
* found as comparands of equijoin predicates in the mergeclause_list.
*
* NB: for mergeclause_list and pathkey structure assumptions, see:
* select_mergejoin_clauses() in joinpath.c
* find_mergeclauses_for_pathkeys() in pathkeys.c
*/
typedef struct
{
PlannerInfo *root;
List *mergeclause_list;
Path *path;
CdbPathLocus locus;
CdbPathLocus *colocus;
bool colocus_eq_locus;
} CdbpathMatchPredsContext;
/*
* A helper function to create a DistributionKey for an EquivalenceClass.
*/
static DistributionKey *
makeDistributionKeyForEC(EquivalenceClass *eclass, Oid opfamily)
{
DistributionKey *dk = makeNode(DistributionKey);
Assert(OidIsValid(opfamily));
dk->dk_eclasses = list_make1(eclass);
dk->dk_opfamily = opfamily;
return dk;
}
/*
* cdbpath_eclass_constant_is_hashable
*
* Iterates through a list of equivalence class members and determines if
* expression in pseudoconstant is hashable under the given hash opfamily.
*
* If there are multiple constants in the equivalence class, it is sufficient
* that one of them is usable.
*/
static bool
cdbpath_eclass_constant_is_hashable(EquivalenceClass *ec, Oid hashOpFamily)
{
ListCell *j;
foreach(j, ec->ec_members)
{
EquivalenceMember *em = (EquivalenceMember *) lfirst(j);
if (!em->em_is_const)
continue;
if (get_opfamily_member(hashOpFamily, em->em_datatype, em->em_datatype,
HTEqualStrategyNumber))
return true;
}
return false;
}
static bool
cdbpath_match_preds_to_distkey_tail(CdbpathMatchPredsContext *ctx,
ListCell *distkeycell)
{
DistributionKey *distkey = (DistributionKey *) lfirst(distkeycell);
DistributionKey *codistkey;
ListCell *cell;
ListCell *rcell;
Assert(CdbPathLocus_IsHashed(ctx->locus) ||
CdbPathLocus_IsHashedOJ(ctx->locus));
/*----------------
* Is there a "<distkey item> = <constant expr>" predicate?
*
* If table T is distributed on cols (C,D,E) and query contains preds
* T.C = U.A AND T.D = <constant expr> AND T.E = U.B
* then we would like to report a match and return the colocus
* (U.A, <constant expr>, U.B)
* so the caller can join T and U by redistributing only U.
* (Note that "T.D = <constant expr>" won't be in the mergeclause_list
* because it isn't a join pred.)
*----------------
*/
codistkey = NULL;
foreach(cell, distkey->dk_eclasses)
{
EquivalenceClass *ec = (EquivalenceClass *) lfirst(cell);
if (CdbEquivClassIsConstant(ec) &&
cdbpath_eclass_constant_is_hashable(ec, distkey->dk_opfamily))
{
codistkey = distkey;
break;
}
}
/* Look for an equijoin comparison to the distkey item. */
if (!codistkey)
{
foreach(rcell, ctx->mergeclause_list)
{
EquivalenceClass *a_ec; /* Corresponding to ctx->path. */
EquivalenceClass *b_ec;
ListCell *i;
RestrictInfo *rinfo = (RestrictInfo *) lfirst(rcell);
update_mergeclause_eclasses(ctx->root, rinfo);
if (bms_is_subset(rinfo->right_relids, ctx->path->parent->relids))
{
a_ec = rinfo->right_ec;
b_ec = rinfo->left_ec;
}
else
{
a_ec = rinfo->left_ec;
b_ec = rinfo->right_ec;
Assert(bms_is_subset(rinfo->left_relids, ctx->path->parent->relids));
}
foreach(i, distkey->dk_eclasses)
{
EquivalenceClass *dk_eclass = (EquivalenceClass *) lfirst(i);
if (dk_eclass == a_ec)
codistkey = makeDistributionKeyForEC(b_ec, distkey->dk_opfamily); /* break earlier? */
}
if (codistkey)
break;
}
}
/* Fail if didn't find a match for this distkey item. */
if (!codistkey)
return false;
/* Might need to build co-locus if locus is outer join source or result. */
if (codistkey != lfirst(distkeycell))
ctx->colocus_eq_locus = false;
/* Match remaining partkey items. */
distkeycell = lnext(distkeycell);
if (distkeycell)
{
if (!cdbpath_match_preds_to_distkey_tail(ctx, distkeycell))
return false;
}
/* Success! Matched all items. Return co-locus if requested. */
if (ctx->colocus)
{
if (ctx->colocus_eq_locus)
*ctx->colocus = ctx->locus;
else if (!distkeycell)
CdbPathLocus_MakeHashed(ctx->colocus, list_make1(codistkey),
CdbPathLocus_NumSegments(ctx->locus));
else
{
ctx->colocus->distkey = lcons(codistkey, ctx->colocus->distkey);
Assert(cdbpathlocus_is_valid(*ctx->colocus));
}
}
return true;
} /* cdbpath_match_preds_to_partkey_tail */
/*
* cdbpath_match_preds_to_partkey
*
* Returns true if an equijoin predicate is found in the mergeclause_list
* for each item of the locus's partitioning key.
*
* (Also, a partkey item that is equal to a constant is treated as a match.)
*
* Readers may refer also to these related functions:
* select_mergejoin_clauses() in joinpath.c
* find_mergeclauses_for_pathkeys() in pathkeys.c
*/
static bool
cdbpath_match_preds_to_distkey(PlannerInfo *root,
List *mergeclause_list,
Path *path,
CdbPathLocus locus,
CdbPathLocus *colocus) /* OUT */
{
CdbpathMatchPredsContext ctx;
if (!CdbPathLocus_IsHashed(locus) &&
!CdbPathLocus_IsHashedOJ(locus))
return false;
Assert(cdbpathlocus_is_valid(locus));
ctx.root = root;
ctx.mergeclause_list = mergeclause_list;
ctx.path = path;
ctx.locus = locus;
ctx.colocus = colocus;
ctx.colocus_eq_locus = true;
return cdbpath_match_preds_to_distkey_tail(&ctx, list_head(locus.distkey));
}
/*
* cdbpath_match_preds_to_both_distkeys
*
* Returns true if the mergeclause_list contains equijoin
* predicates between each item of the outer_locus distkey and
* the corresponding item of the inner_locus distkey.
*
* Readers may refer also to these related functions:
* select_mergejoin_clauses() in joinpath.c
* find_mergeclauses_for_pathkeys() in pathkeys.c
*/
static bool
cdbpath_match_preds_to_both_distkeys(PlannerInfo *root,
List *mergeclause_list,
CdbPathLocus outer_locus,
CdbPathLocus inner_locus)
{
ListCell *outercell;
ListCell *innercell;
List *outer_distkey;
List *inner_distkey;
if (!mergeclause_list ||
CdbPathLocus_NumSegments(outer_locus) != CdbPathLocus_NumSegments(inner_locus) ||
outer_locus.distkey == NIL || inner_locus.distkey == NIL ||
list_length(outer_locus.distkey) != list_length(inner_locus.distkey))
return false;
Assert(CdbPathLocus_IsHashed(outer_locus) ||
CdbPathLocus_IsHashedOJ(outer_locus));
Assert(CdbPathLocus_IsHashed(inner_locus) ||
CdbPathLocus_IsHashedOJ(inner_locus));
outer_distkey = outer_locus.distkey;
inner_distkey = inner_locus.distkey;
forboth(outercell, outer_distkey, innercell, inner_distkey)
{
DistributionKey *outer_dk = (DistributionKey *) lfirst(outercell);
DistributionKey *inner_dk = (DistributionKey *) lfirst(innercell);
ListCell *rcell;
if (outer_dk->dk_opfamily != inner_dk->dk_opfamily)
return false; /* incompatible hashing scheme */
foreach(rcell, mergeclause_list)
{
bool not_found = false;
RestrictInfo *rinfo = (RestrictInfo *) lfirst(rcell);
ListCell *i;
update_mergeclause_eclasses(root, rinfo);
/* Skip predicate if neither side matches outer distkey item. */
foreach(i, outer_dk->dk_eclasses)
{
EquivalenceClass *outer_ec = (EquivalenceClass *) lfirst(i);
if (outer_ec != rinfo->left_ec && outer_ec != rinfo->right_ec)
{
not_found = true;
break;
}
}
if (not_found)
continue;
/* Skip predicate if neither side matches inner distkey item. */
if (inner_dk != outer_dk)
{
ListCell *i;
foreach(i, inner_dk->dk_eclasses)
{
EquivalenceClass *inner_ec = (EquivalenceClass *) lfirst(i);
if (inner_ec != rinfo->left_ec && inner_ec != rinfo->right_ec)
{
not_found = true;
break;
}
}
if (not_found)
continue;
}
/* Found equijoin between outer distkey item & inner distkey item */
break;
}
/* Fail if didn't find equijoin between this pair of distkey items. */
if (!rcell)
return false;
}
return true;
} /* cdbpath_match_preds_to_both_distkeys */
/*
* cdbpath_distkeys_from_preds
*
* Makes a CdbPathLocus for repartitioning, driven by
* the equijoin predicates in the mergeclause_list (a List of RestrictInfo).
* Returns true if successful, or false if no usable equijoin predicates.
*
* Readers may refer also to these related functions:
* select_mergejoin_clauses() in joinpath.c
* make_pathkeys_for_mergeclauses() in pathkeys.c
*/
static bool
cdbpath_distkeys_from_preds(PlannerInfo *root,
List *mergeclause_list,
Path *a_path,
int numsegments,
CdbPathLocus *a_locus, /* OUT */
CdbPathLocus *b_locus) /* OUT */
{
List *a_distkeys = NIL;
List *b_distkeys = NIL;
ListCell *rcell;
foreach(rcell, mergeclause_list)
{
RestrictInfo *rinfo = (RestrictInfo *) lfirst(rcell);
Oid lhs_opno;
Oid rhs_opno;
Oid opfamily;
update_mergeclause_eclasses(root, rinfo);
/*
* skip non-hashable keys
*/
if (!rinfo->hashjoinoperator)
continue;
/*
* look up a hash operator family that is compatible for the left and right datatypes
* of the hashjoin = operator
*/
if (!get_compatible_hash_operators_and_family(rinfo->hashjoinoperator,
&lhs_opno, &rhs_opno, &opfamily))
continue;
/* Left & right pathkeys are usually the same... */
if (!b_distkeys && rinfo->left_ec == rinfo->right_ec)
{
ListCell *i;
bool found = false;
foreach(i, a_distkeys)
{
DistributionKey *distkey = (DistributionKey *) lfirst(i);
EquivalenceClass *dk_eclass;
/*
* we only create Hashed DistributionKeys with a single eclass
* in this function.
*/
Assert(list_length(distkey->dk_eclasses) == 1);
dk_eclass = (EquivalenceClass *) linitial(distkey->dk_eclasses);
if (dk_eclass == rinfo->left_ec)
{
found = true;
break;
}
}
if (!found)
{
DistributionKey *a_dk = makeDistributionKeyForEC(rinfo->left_ec, opfamily);
a_distkeys = lappend(a_distkeys, a_dk);
}
}
/* ... except in outer join ON-clause. */
else
{
EquivalenceClass *a_ec;
EquivalenceClass *b_ec;
ListCell *i;
bool found = false;
if (bms_is_subset(rinfo->right_relids, a_path->parent->relids))
{
a_ec = rinfo->right_ec;
b_ec = rinfo->left_ec;
}
else
{
a_ec = rinfo->left_ec;
b_ec = rinfo->right_ec;
Assert(bms_is_subset(rinfo->left_relids, a_path->parent->relids));
}
if (!b_ec)
b_ec = a_ec;
/*
* Convoluted logic to ensure that (a_ec not in a_distkeys) AND
* (b_ec not in b_distkeys)
*/
found = false;
foreach(i, a_distkeys)
{
DistributionKey *distkey = (DistributionKey *) lfirst(i);
EquivalenceClass *dk_eclass;
/*
* we only create Hashed DistributionKeys with a single eclass
* in this function.
*/
Assert(list_length(distkey->dk_eclasses) == 1);
dk_eclass = (EquivalenceClass *) linitial(distkey->dk_eclasses);
if (dk_eclass == a_ec)
{
found = true;
break;
}
}
if (!found)
{
foreach(i, b_distkeys)
{
DistributionKey *distkey = (DistributionKey *) lfirst(i);
EquivalenceClass *dk_eclass;
/*
* we only create Hashed DistributionKeys with a single eclass
* in this function.
*/
Assert(list_length(distkey->dk_eclasses) == 1);
dk_eclass = (EquivalenceClass *) linitial(distkey->dk_eclasses);
if (dk_eclass == b_ec)
{
found = true;
break;
}
}
}
if (!found)
{
DistributionKey *a_dk = makeDistributionKeyForEC(a_ec, opfamily);
DistributionKey *b_dk = makeDistributionKeyForEC(b_ec, opfamily);
a_distkeys = lappend(a_distkeys, a_dk);
b_distkeys = lappend(b_distkeys, b_dk);
}
}
if (list_length(a_distkeys) >= 20)
break;
}
if (!a_distkeys)
return false;
CdbPathLocus_MakeHashed(a_locus, a_distkeys, numsegments);
if (b_distkeys)
CdbPathLocus_MakeHashed(b_locus, b_distkeys, numsegments);
else
*b_locus = *a_locus;
return true;
} /* cdbpath_distkeys_from_preds */
/*
* Add a RowIdExpr to the target list of 'path'. Returns the ID
* of the generated rowid expression in *rowidexpr_id.
*/
static Path *
add_rowid_to_path(PlannerInfo *root, Path *path, int *rowidexpr_id)
{
RowIdExpr *rowidexpr;
PathTarget *newpathtarget;
/*
* 'last_rowidexpr_id' is used to generate a unique ID for the RowIdExpr
* node that we generate. It only needs to be unique within this query
* plan, and the simplest way to achieve that is to just have a global
* counter. (Actually, it isn't really needed at the moment because the
* deduplication is always done immediately on top of the join, so two
* different RowIdExprs should never appear in the same part of the plan
* tree. But it might come handy when debugging, if nothing else.
* XXX: If we start to rely on it for something important, consider
* overflow behavior more carefully.)
*/
static uint32 last_rowidexpr_id = 0;
rowidexpr = makeNode(RowIdExpr);
last_rowidexpr_id++;
*rowidexpr_id = rowidexpr->rowidexpr_id = (int) last_rowidexpr_id;
newpathtarget = copy_pathtarget(path->pathtarget);
add_column_to_pathtarget(newpathtarget, (Expr *) rowidexpr, 0);
return (Path *) create_projection_path_with_quals(root, path->parent,
path, newpathtarget,
NIL, true);
}
/*
* cdbpath_motion_for_join
*
* Decides where a join should be done. Adds Motion operators atop
* the subpaths if needed to deliver their results to the join locus.
* Returns the join locus if ok, or a null locus otherwise. If
* jointype is JOIN_SEMI_DEDUP or JOIN_SEMI_DEDUP_REVERSE, this also
* tacks a RowIdExpr on one side of the join, and *p_rowidexpr_id is
* set to the ID of that. The caller is expected to uniquefy
* the result after the join, passing the rowidexpr_id to
* create_unique_rowid_path().
*
* mergeclause_list is a List of RestrictInfo. Its members are
* the equijoin predicates between the outer and inner rel.
* It comes from select_mergejoin_clauses() in joinpath.c.
*/
CdbPathLocus
cdbpath_motion_for_join(PlannerInfo *root,
JoinType jointype, /* JOIN_INNER/FULL/LEFT/RIGHT/IN */
Path **p_outer_path, /* INOUT */
Path **p_inner_path, /* INOUT */
int *p_rowidexpr_id, /* OUT */
List *redistribution_clauses, /* equijoin RestrictInfo list */
List *restrict_clauses,
List *outer_pathkeys,
List *inner_pathkeys,
bool outer_require_existing_order,
bool inner_require_existing_order)
{
CdbpathMfjRel outer;
CdbpathMfjRel inner;
int numsegments;
bool join_quals_contain_outer_references;
ListCell *lc;
*p_rowidexpr_id = 0;
outer.pathkeys = outer_pathkeys;
inner.pathkeys = inner_pathkeys;
outer.path = *p_outer_path;
inner.path = *p_inner_path;
outer.locus = outer.path->locus;
inner.locus = inner.path->locus;
CdbPathLocus_MakeNull(&outer.move_to);
CdbPathLocus_MakeNull(&inner.move_to);
Assert(cdbpathlocus_is_valid(outer.locus));
Assert(cdbpathlocus_is_valid(inner.locus));
/*
* Does the join quals contain references to outer query? If so, we must
* evaluate them in the outer query's locus. That means pulling both
* inputs to outer locus, and performing the join there.
*
* XXX: If there are pseudoconstant quals, they will be executed by a
* gating Result with a One-Time Filter. In that case, the join's inputs
* wouldn't need to be brought to the outer locus. We could execute the
* join normally, and bring the result to the outer locus and put the
* gating Result above the Motion, instead. But for now, we're not smart
* like that.
*/
join_quals_contain_outer_references = false;
foreach(lc, restrict_clauses)
{
RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
if (rinfo->contain_outer_query_references)
{
join_quals_contain_outer_references = true;
break;
}
}
/*
* Locus type Replicated can only be generated by join operation.
* And in the function cdbpathlocus_join there is a rule:
* <any locus type> join <Replicated> => any locus type
* Proof by contradiction, it shows that when code arrives here,
* is is impossible that any of the two input paths' locus
* is Replicated. So we add two asserts here.
*/
Assert(!CdbPathLocus_IsReplicated(outer.locus));
Assert(!CdbPathLocus_IsReplicated(inner.locus));
if (CdbPathLocus_IsReplicated(outer.locus) ||
CdbPathLocus_IsReplicated(inner.locus))
goto fail;
outer.has_wts = cdbpath_contains_wts(outer.path);
inner.has_wts = cdbpath_contains_wts(inner.path);
/* For now, inner path should not contain WorkTableScan */
Assert(!inner.has_wts);
/*
* If outer rel contains WorkTableScan and inner rel is hash distributed,
* unfortunately we have to pretend that inner is randomly distributed,
* otherwise we may end up with redistributing outer rel.
*/
if (outer.has_wts && inner.locus.distkey != NIL)
CdbPathLocus_MakeStrewn(&inner.locus,
CdbPathLocus_NumSegments(inner.locus));
/*
* Caller can specify an ordering for each source path that is the same as
* or weaker than the path's existing ordering. Caller may insist that we
* do not add motion that would lose the specified ordering property;
* otherwise the given ordering is preferred but not required. A required
* NIL ordering means no motion is allowed for that path.
*/
outer.require_existing_order = outer_require_existing_order;
inner.require_existing_order = inner_require_existing_order;
/*
* Don't consider replicating the preserved rel of an outer join, or the
* current-query rel of a join between current query and subquery.
*
* Path that contains WorkTableScan cannot be replicated.
*/
/* ok_to_replicate means broadcast */
outer.ok_to_replicate = !outer.has_wts;
inner.ok_to_replicate = true;
switch (jointype)
{
case JOIN_INNER:
case JOIN_UNIQUE_OUTER:
case JOIN_UNIQUE_INNER:
break;
case JOIN_SEMI:
case JOIN_ANTI:
case JOIN_LEFT:
case JOIN_LASJ_NOTIN:
outer.ok_to_replicate = false;
break;
case JOIN_RIGHT:
inner.ok_to_replicate = false;
break;
case JOIN_FULL:
outer.ok_to_replicate = false;
inner.ok_to_replicate = false;
break;
case JOIN_DEDUP_SEMI:
/*
* In this plan type, we generate a unique row ID on the outer
* side of the join, perform the join, possibly broadcasting the
* outer side, and remove duplicates after the join, so that only
* one row for each input outer row remains.
*
* If the outer input is General or SegmentGeneral, it's available
* in all the segments, but we cannot reliably generate a row ID
* to distinguish each logical row in that case. So force the
* input to a single node first in that case.
*
* In previous Greenplum versions, we assumed that we can generate
* a unique row ID for General paths, by generating the same
* sequence of numbers on each segment. That works as long as the
* rows are in the same order on each segment, but it seemed like
* a risky assumption. And it didn't work on SegmentGeneral paths
* (i.e. replicated tables) anyway.
*/
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
goto fail;
inner.ok_to_replicate = false;
outer.path = add_rowid_to_path(root, outer.path, p_rowidexpr_id);
*p_outer_path = outer.path;
break;
case JOIN_DEDUP_SEMI_REVERSE:
/* 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
goto fail;
outer.ok_to_replicate = false;
inner.path = add_rowid_to_path(root, inner.path, p_rowidexpr_id);
*p_inner_path = inner.path;
break;
default:
/*
* The caller should already have transformed
* JOIN_UNIQUE_INNER/OUTER into JOIN_INNER
*/
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;
if (join_quals_contain_outer_references)
{
if (CdbPathLocus_IsOuterQuery(outer.locus) &&
CdbPathLocus_IsOuterQuery(inner.locus))
return outer.locus;
if (!CdbPathLocus_IsOuterQuery(outer.locus))
CdbPathLocus_MakeOuterQuery(&outer.move_to);
if (!CdbPathLocus_IsOuterQuery(inner.locus))
CdbPathLocus_MakeOuterQuery(&inner.move_to);
}
else if (CdbPathLocus_IsOuterQuery(outer.locus) ||
CdbPathLocus_IsOuterQuery(inner.locus))
{
/*
* If one side of the join has "outer query" locus, must bring the
* other side there too.
*/
if (CdbPathLocus_IsOuterQuery(outer.locus) &&
CdbPathLocus_IsOuterQuery(inner.locus))
return outer.locus;
if (CdbPathLocus_IsOuterQuery(outer.locus))
inner.move_to = outer.locus;
else
outer.move_to = inner.locus;
}
else if (CdbPathLocus_IsGeneral(outer.locus) ||
CdbPathLocus_IsGeneral(inner.locus))
{
/*
* Motion not needed if either source is everywhere (e.g. a constant).
*
* But if a row is everywhere and is preserved in an outer join, we don't
* want to preserve it in every qExec process where it is unmatched,
* because that would produce duplicate null-augmented rows. So in that
* case, bring all the partitions to a single qExec to be joined. CDB
* TODO: Can this case be handled without introducing a bottleneck?
*/
/*
* The logic for the join result's locus is (outer's locus is general):
* 1. if outer is ok to replicated, then result's locus is the same
* as inner's locus
* 2. if outer is not ok to replicated (like left join or wts cases)
* 2.1 if inner's locus is hashed or hashOJ, we try to redistribute
* outer as the inner, if fails, make inner singleQE
* 2.2 if inner's locus is strewn, we try to redistribute
* outer and inner, if fails, make inner singleQE
* 2.3 just return the inner's locus, no motion is needed
*/
CdbpathMfjRel *general = &outer;
CdbpathMfjRel *other = &inner;
/*
* both are general, the result is general
*/
if (CdbPathLocus_IsGeneral(outer.locus) &&
CdbPathLocus_IsGeneral(inner.locus))
return outer.locus;
if (CdbPathLocus_IsGeneral(inner.locus))
{
general = &inner;
other = &outer;
}
/* numsegments of General locus is always -1 */
Assert(CdbPathLocus_NumSegments(general->locus) == -1);
/*
* If general can happen everywhere (ok_to_replicate)
* then it acts like identity:
* General join other_locus => other_locus
*/
if (general->ok_to_replicate)
return other->locus;
if (!CdbPathLocus_IsPartitioned(other->locus))
{
/*
* If general is not ok_to_replicate, for example,
* generate_series(1, 10) left join xxxx, only for
* some specific locus types general can act as
* identity:
* General join other_locus => other_locus, if and only if
* other_locus in (singleQE, Entry).
* Here other's locus:
* - cannot be general (it has already handled)
* - cannot be replicated (assert at the beginning of the function)
*/
Assert(CdbPathLocus_IsBottleneck(other->locus) ||
CdbPathLocus_IsSegmentGeneral(other->locus));
return other->locus;
}
/*
* If other's locus is partitioned, we first try to
* add redistribute motion, if fails, we gather other
* to singleQE.
*/
else if (!try_redistribute(root, general, other, redistribution_clauses))
{
/*
* FIXME: do we need test other's movable?
*/
CdbPathLocus_MakeSingleQE(&other->move_to,
CdbPathLocus_NumSegments(other->locus));
}
}
else if (CdbPathLocus_IsSegmentGeneral(outer.locus) ||
CdbPathLocus_IsSegmentGeneral(inner.locus))
{
/*
* the whole branch handles the case that at least
* one of the two locus is SegmentGeneral. The logic
* is:
* - if both are SegmentGeneral:
* 1. if both locus are equal, no motion needed, simply return
* 2. For update cases. If resultrelation
* is SegmentGeneral, the update must execute
* on each segment of the resultrelation, if resultrelation's
* numsegments is larger, the only solution is to broadcast
* other
* 3. no motion is needed, change both numsegments to common
* - if only one of them is SegmentGeneral :
* 1. consider update case, if resultrelation is SegmentGeneral,
* the only solution is to broadcast the other
* 2. if other's locus is singleQE or entry, make SegmentGeneral
* to other's locus
* 3. the remaining possibility of other's locus is partitioned
* 3.1 if SegmentGeneral is not ok_to_replicate, try to
* add redistribute motion, if fails gather each to
* singleQE
* 3.2 if SegmentGeneral's numsegments is larger, just return
* other's locus
* 3.3 try to add redistribute motion, if fails, gather each
* to singleQE
*/
CdbpathMfjRel *segGeneral;
CdbpathMfjRel *other;
if (CdbPathLocus_IsSegmentGeneral(outer.locus) &&
CdbPathLocus_IsSegmentGeneral(inner.locus))
{
/*
* use_common to indicate whether we should
* return a segmentgeneral locus with common
* numsegments.
*/
bool use_common = true;
/*
* Handle the case two same locus
*/
if (CdbPathLocus_NumSegments(outer.locus) ==
CdbPathLocus_NumSegments(inner.locus))
return inner.locus;
/*
* Now, two locus' numsegments not equal
* We should consider update resultrelation
* if update,
* - resultrelation's numsegments larger, then
* we should broadcast the other
* - otherwise, results is common
* else:
* common
*/
if (root->upd_del_replicated_table > 0)
{
if ((CdbPathLocus_NumSegments(outer.locus) >
CdbPathLocus_NumSegments(inner.locus)) &&
bms_is_member(root->upd_del_replicated_table,
outer.path->parent->relids))
{
/*
* the updated resultrelation is replicated table
* and its numsegments is larger, we should broadcast
* the other path
*/
if (!inner.ok_to_replicate)
goto fail;
/*
* FIXME: do we need to test inner's movable?
*/
CdbPathLocus_MakeReplicated(&inner.move_to,
CdbPathLocus_NumSegments(outer.locus));
use_common = false;
}
else if ((CdbPathLocus_NumSegments(outer.locus) <
CdbPathLocus_NumSegments(inner.locus)) &&
bms_is_member(root->upd_del_replicated_table,
inner.path->parent->relids))
{
/*
* the updated resultrelation is replicated table
* and its numsegments is larger, we should broadcast
* the other path
*/
if (!outer.ok_to_replicate)
goto fail;
/*
* FIXME: do we need to test outer's movable?
*/
CdbPathLocus_MakeReplicated(&outer.move_to,
CdbPathLocus_NumSegments(inner.locus));
use_common = false;
}
}
if (use_common)
{
/*
* The statement is not update a replicated table.
* Just return the segmentgeneral with a smaller numsegments.
*/
numsegments = CdbPathLocus_CommonSegments(inner.locus,
outer.locus);
outer.locus.numsegments = numsegments;
inner.locus.numsegments = numsegments;
return inner.locus;
}
}
else
{
if (CdbPathLocus_IsSegmentGeneral(outer.locus))
{
segGeneral = &outer;
other = &inner;
}
else
{
segGeneral = &inner;
other = &outer;
}
Assert(CdbPathLocus_IsBottleneck(other->locus) ||
CdbPathLocus_IsPartitioned(other->locus));
/*
* For UPDATE/DELETE, replicated table can't guarantee a logic row has
* same ctid or item pointer on each copy. If we broadcast matched tuples
* to all segments, the segments may update the wrong tuples or can't
* find a valid tuple according to ctid or item pointer.
*
* So For UPDATE/DELETE on replicated table, we broadcast other path so
* all target tuples can be selected on all copys and then be updated
* locally.
*/
if (root->upd_del_replicated_table > 0 &&
bms_is_member(root->upd_del_replicated_table,
segGeneral->path->parent->relids))
{
/*
* For UPDATE on a replicated table, we have to do it
* everywhere so that for each segment, we have to collect
* all the information of other that is we should broadcast it
*/
/*
* FIXME: do we need to test other's movable?
*/
CdbPathLocus_MakeReplicated(&other->move_to,
CdbPathLocus_NumSegments(segGeneral->locus));
}
else if (CdbPathLocus_IsBottleneck(other->locus))
{
/*
* if the locus type is equal and segment count is unequal,
* we will dispatch the one on more segments to the other
*/
numsegments = CdbPathLocus_CommonSegments(segGeneral->locus,
other->locus);
segGeneral->move_to = other->locus;
segGeneral->move_to.numsegments = numsegments;
}
else
{
/*
* This branch handles for partitioned other locus
* hashed, hashoj, strewn
*/
Assert(CdbPathLocus_IsPartitioned(other->locus));
if (!segGeneral->ok_to_replicate)
{
if (!try_redistribute(root, segGeneral,
other, redistribution_clauses))
{
/*
* FIXME: do we need to test movable?
*/
CdbPathLocus_MakeSingleQE(&segGeneral->move_to,
CdbPathLocus_NumSegments(segGeneral->locus));
CdbPathLocus_MakeSingleQE(&other->move_to,
CdbPathLocus_NumSegments(other->locus));
}
}
else
{
/*
* If all other's segments have segGeneral stored, then no motion
* is needed.
*
* A sql to reach here:
* select * from d2 a join r1 b using (c1);
* where d2 is a replicated table on 2 segment,
* r1 is a random table on 1 segments.
*/
if (CdbPathLocus_NumSegments(segGeneral->locus) >=
CdbPathLocus_NumSegments(other->locus))
return other->locus;
else
{
if (!try_redistribute(root, segGeneral,
other, redistribution_clauses))
{
numsegments = CdbPathLocus_CommonSegments(segGeneral->locus,
other->locus);
/*
* FIXME: do we need to test movable?
*/
CdbPathLocus_MakeSingleQE(&segGeneral->move_to, numsegments);
CdbPathLocus_MakeSingleQE(&other->move_to, numsegments);
}
}
}
}
}
}
/*
* Is either source confined to a single process? NB: Motion to a single
* process (qDisp or qExec) is the only motion in which we may use Merge
* Receive to preserve an existing ordering.
*/
else if (CdbPathLocus_IsBottleneck(outer.locus) ||
CdbPathLocus_IsBottleneck(inner.locus))
{ /* singleQE or entry db */
CdbpathMfjRel *single = &outer;
CdbpathMfjRel *other = &inner;
bool single_immovable = (outer.require_existing_order &&
!outer_pathkeys) || outer.has_wts;
bool other_immovable = inner.require_existing_order &&
!inner_pathkeys;
/*
* If each of the sources has a single-process locus, then assign both
* sources and the join to run in the same process, without motion.
* The slice will be run on the entry db if either source requires it.
*/
if (CdbPathLocus_IsEntry(single->locus))
{
if (CdbPathLocus_IsBottleneck(other->locus))
return single->locus;
}
else if (CdbPathLocus_IsSingleQE(single->locus))
{
if (CdbPathLocus_IsBottleneck(other->locus))
{
/*
* Can join directly on one of the common segments.
*/
numsegments = CdbPathLocus_CommonSegments(outer.locus,
inner.locus);
other->locus.numsegments = numsegments;
return other->locus;
}
}
/* Let 'single' be the source whose locus is singleQE or entry. */
else
{
CdbSwap(CdbpathMfjRel *, single, other);
CdbSwap(bool, single_immovable, other_immovable);
}
Assert(CdbPathLocus_IsBottleneck(single->locus));
Assert(CdbPathLocus_IsPartitioned(other->locus));
/* If the bottlenecked rel can't be moved, bring the other rel to it. */
if (single_immovable)
{
if (other_immovable)
goto fail;
else
other->move_to = single->locus;
}
/* Redistribute single rel if joining on other rel's partitioning key */
else if (cdbpath_match_preds_to_distkey(root,
redistribution_clauses,
other->path,
other->locus,
&single->move_to)) /* OUT */
{
AssertEquivalent(CdbPathLocus_NumSegments(other->locus),
CdbPathLocus_NumSegments(single->move_to));
}
/* Replicate single rel if cheaper than redistributing both rels. */
else if (single->ok_to_replicate &&
(single->bytes * CdbPathLocus_NumSegments(other->locus) <
single->bytes + other->bytes))
CdbPathLocus_MakeReplicated(&single->move_to,
CdbPathLocus_NumSegments(other->locus));
/*
* Redistribute both rels on equijoin cols.
*
* Redistribute both to the same segments, here we choose the
* same segments with other.
*/
else if (!other_immovable &&
cdbpath_distkeys_from_preds(root,
redistribution_clauses,
single->path,
CdbPathLocus_NumSegments(other->locus),
&single->move_to, /* OUT */
&other->move_to)) /* OUT */
{
/* ok */
}
/* Broadcast single rel for below cases. */
else if (single->ok_to_replicate &&
(other_immovable ||
single->bytes < other->bytes ||
other->has_wts))
CdbPathLocus_MakeReplicated(&single->move_to,
CdbPathLocus_NumSegments(other->locus));
/* Last resort: If possible, move all partitions of other rel to single QE. */
else if (!other_immovable)
other->move_to = single->locus;
else
goto fail;
} /* singleQE or entry */
/*
* No motion if partitioned alike and joining on the partitioning keys.
*/
else if (cdbpath_match_preds_to_both_distkeys(root, redistribution_clauses,
outer.locus, inner.locus))
return cdbpathlocus_join(jointype, outer.locus, inner.locus);
/*
* Both sources are partitioned. Redistribute or replicate one or both.
*/
else
{ /* partitioned */
CdbpathMfjRel *large_rel = &outer;
CdbpathMfjRel *small_rel = &inner;
/* Which rel is bigger? */
if (large_rel->bytes < small_rel->bytes)
CdbSwap(CdbpathMfjRel *, large_rel, small_rel);
/* Both side are distribued in 1 segment, it can join without motion. */
if (CdbPathLocus_NumSegments(large_rel->locus) == 1 &&
CdbPathLocus_NumSegments(small_rel->locus) == 1)
return large_rel->locus;
/* If joining on larger rel's partitioning key, redistribute smaller. */
if (!small_rel->require_existing_order &&
cdbpath_match_preds_to_distkey(root,
redistribution_clauses,
large_rel->path,
large_rel->locus,
&small_rel->move_to)) /* OUT */
{
AssertEquivalent(CdbPathLocus_NumSegments(large_rel->locus),
CdbPathLocus_NumSegments(small_rel->move_to));
}
/*
* Replicate smaller rel if cheaper than redistributing larger rel.
* But don't replicate a rel that is to be preserved in outer join.
*/
else if (!small_rel->require_existing_order &&
small_rel->ok_to_replicate &&
(small_rel->bytes * CdbPathLocus_NumSegments(large_rel->locus) <
large_rel->bytes))
CdbPathLocus_MakeReplicated(&small_rel->move_to,
CdbPathLocus_NumSegments(large_rel->locus));
/*
* Replicate larger rel if cheaper than redistributing smaller rel.
* But don't replicate a rel that is to be preserved in outer join.
*/
else if (!large_rel->require_existing_order &&
large_rel->ok_to_replicate &&
(large_rel->bytes * CdbPathLocus_NumSegments(small_rel->locus) <
small_rel->bytes))
CdbPathLocus_MakeReplicated(&large_rel->move_to,
CdbPathLocus_NumSegments(small_rel->locus));
/* If joining on smaller rel's partitioning key, redistribute larger. */
else if (!large_rel->require_existing_order &&
cdbpath_match_preds_to_distkey(root,
redistribution_clauses,
small_rel->path,
small_rel->locus,
&large_rel->move_to)) /* OUT */
{
AssertEquivalent(CdbPathLocus_NumSegments(small_rel->locus),
CdbPathLocus_NumSegments(large_rel->move_to));
}
/* Replicate smaller rel if cheaper than redistributing both rels. */
else if (!small_rel->require_existing_order &&
small_rel->ok_to_replicate &&
(small_rel->bytes * CdbPathLocus_NumSegments(large_rel->locus) <
small_rel->bytes + large_rel->bytes))
CdbPathLocus_MakeReplicated(&small_rel->move_to,
CdbPathLocus_NumSegments(large_rel->locus));
/* Replicate largeer rel if cheaper than redistributing both rels. */
else if (!large_rel->require_existing_order &&
large_rel->ok_to_replicate &&
(large_rel->bytes * CdbPathLocus_NumSegments(small_rel->locus) <
large_rel->bytes + small_rel->bytes))
CdbPathLocus_MakeReplicated(&large_rel->move_to,
CdbPathLocus_NumSegments(small_rel->locus));
/*
* Redistribute both rels on equijoin cols.
*
* the two results should all be distributed on the same segments,
* here we make them the same with common segments for safe
* TODO: how about distribute them both to ALL segments?
*/
else if (!small_rel->require_existing_order &&
!small_rel->has_wts &&
!large_rel->require_existing_order &&
!large_rel->has_wts &&
cdbpath_distkeys_from_preds(root,
redistribution_clauses,
large_rel->path,
CdbPathLocus_CommonSegments(large_rel->locus,
small_rel->locus),
&large_rel->move_to,
&small_rel->move_to))
{
/* ok */
}
/*
* No usable equijoin preds, or couldn't consider the preferred
* motion. Replicate one rel if possible. MPP TODO: Consider number of
* seg dbs per host.
*/
else if (!small_rel->require_existing_order &&
small_rel->ok_to_replicate)
CdbPathLocus_MakeReplicated(&small_rel->move_to,
CdbPathLocus_NumSegments(large_rel->locus));
else if (!large_rel->require_existing_order &&
large_rel->ok_to_replicate)
CdbPathLocus_MakeReplicated(&large_rel->move_to,
CdbPathLocus_NumSegments(small_rel->locus));
/* Last resort: Move both rels to a single qExec. */
else
{
int numsegments = CdbPathLocus_CommonSegments(outer.locus,
inner.locus);
CdbPathLocus_MakeSingleQE(&outer.move_to, numsegments);
CdbPathLocus_MakeSingleQE(&inner.move_to, numsegments);
}
} /* partitioned */
/*
* Move outer.
*/
if (!CdbPathLocus_IsNull(outer.move_to))
{
outer.path = cdbpath_create_motion_path(root,
outer.path,
outer_pathkeys,
outer.require_existing_order,
outer.move_to);
if (!outer.path) /* fail if outer motion not feasible */
goto fail;
if (IsA(outer.path, MaterialPath) && !root->config->may_rescan)
{
/*
* If we are the outer path and can never be rescanned,
* we could remove the materialize path.
*/
MaterialPath *mpath = (MaterialPath *) outer.path;
outer.path = mpath->subpath;
}
}
/*
* Move inner.
*/
if (!CdbPathLocus_IsNull(inner.move_to))
{
inner.path = cdbpath_create_motion_path(root,
inner.path,
inner_pathkeys,
inner.require_existing_order,
inner.move_to);
if (!inner.path) /* fail if inner motion not feasible */
goto fail;
}
/*
* Ok to join. Give modified subpaths to caller.
*/
*p_outer_path = outer.path;
*p_inner_path = inner.path;
/* Tell caller where the join will be done. */
return cdbpathlocus_join(jointype, outer.path->locus, inner.path->locus);
fail: /* can't do this join */
CdbPathLocus_MakeNull(&outer.move_to);
return outer.move_to;
} /* cdbpath_motion_for_join */
/*
* Does the path contain WorkTableScan?
*/
bool
cdbpath_contains_wts(Path *path)
{
JoinPath *joinPath;
AppendPath *appendPath;
ListCell *lc;
if (IsJoinPath(path))
{
joinPath = (JoinPath *) path;
if (cdbpath_contains_wts(joinPath->outerjoinpath)
|| cdbpath_contains_wts(joinPath->innerjoinpath))
return true;
else
return false;
}
else if (IsA(path, AppendPath))
{
appendPath = (AppendPath *) path;
foreach(lc, appendPath->subpaths)
{
if (cdbpath_contains_wts((Path *) lfirst(lc)))
return true;
}
return false;
}
return path->pathtype == T_WorkTableScan;
}
/*
* has_redistributable_clause
* If the restrictinfo's clause is redistributable, return true.
*/
bool
has_redistributable_clause(RestrictInfo *restrictinfo)
{
return restrictinfo->hashjoinoperator != InvalidOid;
}
/*
* try_redistribute
* helper function for A join B when
* - A's locus is general or segmentgeneral
* - B's locus is partitioned
* it tries to redistribute A to B's locus
* or redistribute both A and B to the same
* partitioned locus.
*
* return values:
* - true: redistributed motion has been added for A
* - false: cannot add redistributed motion, caller should
* continue to find other solutions.
*/
static bool
try_redistribute(PlannerInfo *root, CdbpathMfjRel *g, CdbpathMfjRel *o,
List *redistribution_clauses)
{
bool g_immovable;
bool o_immovable;
Assert(CdbPathLocus_IsGeneral(g->locus) ||
CdbPathLocus_IsSegmentGeneral(g->locus));
Assert(CdbPathLocus_IsPartitioned(o->locus));
/*
* we cannot add motion if requiring order.
* has_wts can be true only for general locus
* otherwise, it is false and not impact the
* value of <x>_immovable.
*/
g_immovable = (g->require_existing_order &&
!g->pathkeys) || g->has_wts;
/*
* if g cannot be added motion on,
* we should return immediately.
*/
if (g_immovable)
return false;
o_immovable = (o->require_existing_order &&
!o->pathkeys) || o->has_wts;
if (CdbPathLocus_IsHashed(o->locus) ||
CdbPathLocus_IsHashedOJ(o->locus))
{
/*
* first try to only redistribute g as o's locus
* if fails then try to redistribute both g and o
*/
if (cdbpath_match_preds_to_distkey(root,
redistribution_clauses,
o->path,
o->locus,
&g->move_to))
return true;
else
{
/*
* both g and o can be added motion on,
* we should try each possible case.
*/
int numsegments;
if (CdbPathLocus_IsGeneral(g->locus))
numsegments = CdbPathLocus_NumSegments(o->locus);
else
numsegments = CdbPathLocus_CommonSegments(o->locus, g->locus);
if(cdbpath_distkeys_from_preds(root,
redistribution_clauses,
o->path,
numsegments,
&o->move_to,
&g->move_to))
{
return true;
}
}
}
else
{
/*
* the only possible solution is to
* redistributed both g and o, so
* both g and o should be movable.
*/
int numsegments;
if (CdbPathLocus_IsGeneral(g->locus))
numsegments = CdbPathLocus_NumSegments(o->locus);
else
numsegments = CdbPathLocus_CommonSegments(o->locus, g->locus);
if (!o_immovable &&
cdbpath_distkeys_from_preds(root,
redistribution_clauses,
o->path,
numsegments,
&o->move_to,
&g->move_to))
{
return true;
}
}
/*
* fail to redistribute, return false
* to let caller know.
*/
return false;
}
/*
* Add a suitable Motion Path so that the input tuples from 'subpath' are
* distributed correctly for insertion into target table.
*/
Path *
create_motion_path_for_ctas(PlannerInfo *root, GpPolicy *policy, Path *subpath)
{
GpPolicyType policyType = policy->ptype;
if (policyType == POLICYTYPE_PARTITIONED && policy->nattrs == 0)
{
/*
* If the target table is DISTRIBUTED RANDOMLY, and the input data
* is already partitioned, we could let the insertions happen where
* they are. But to ensure more random distribution, redistribute.
* This is different from create_motion_path_for_insert().
*/
CdbPathLocus targetLocus;
CdbPathLocus_MakeStrewn(&targetLocus, policy->numsegments);
return cdbpath_create_motion_path(root, subpath, NIL, false, targetLocus);
}
else
return create_motion_path_for_insert(root, policy, subpath);
}
/*
* Add a suitable Motion Path so that the input tuples from 'subpath' are
* distributed correctly for insertion into target table.
*/
Path *
create_motion_path_for_insert(PlannerInfo *root, GpPolicy *policy,
Path *subpath)
{
GpPolicyType policyType = policy->ptype;
CdbPathLocus targetLocus;
if (policyType == POLICYTYPE_PARTITIONED)
{
/*
* A query to reach here: INSERT INTO t1 VALUES(1).
* There is no need to add a motion from General, we could
* simply put General on the same segments with target table.
*/
/* FIXME: also do this for other targetPolicyType? */
/* FIXME: also do this for all the subplans */
if (CdbPathLocus_IsGeneral(subpath->locus))
{
subpath->locus.numsegments = policy->numsegments;
}
targetLocus = cdbpathlocus_for_insert(root, policy, subpath->pathtarget);
if (policy->nattrs == 0 && CdbPathLocus_IsPartitioned(subpath->locus))
{
/*
* If the target table is DISTRIBUTED RANDOMLY, we can insert the
* rows anywhere. So if the input path is already partitioned, let
* the insertions happen where they are.
*
* If you `explain` the query insert into tab_random select * from tab_partition
* there is not Motion node in plan. However, it is not means that the query only
* execute in entry db. It is dispatched to QE and do everything well as we expect.
*
* But, we need to grant a Motion node if target locus' segnumber is different with
* subpath.
*/
if(targetLocus.numsegments != subpath->locus.numsegments)
{
CdbPathLocus_MakeStrewn(&targetLocus, policy->numsegments);
subpath = cdbpath_create_motion_path(root, subpath, NIL, false, targetLocus);
}
}
else if (CdbPathLocus_IsNull(targetLocus))
{
/* could not create DistributionKeys to represent the distribution keys. */
CdbPathLocus_MakeStrewn(&targetLocus, policy->numsegments);
subpath = (Path *) make_motion_path(root, subpath, targetLocus, false, policy);
}
else
{
subpath = cdbpath_create_motion_path(root, subpath, NIL, false, targetLocus);
}
}
else if (policyType == POLICYTYPE_ENTRY)
{
/*
* Query result needs to be brought back to the QD.
*/
CdbPathLocus_MakeEntry(&targetLocus);
subpath = cdbpath_create_motion_path(root, subpath, NIL, false, targetLocus);
}
else if (policyType == POLICYTYPE_REPLICATED)
{
/* try to optimize insert with no motion introduced into */
if (optimizer_replicated_table_insert &&
!contain_volatile_functions((Node *)subpath->pathtarget->exprs) &&
!contain_volatile_functions((Node *)root->parse->havingQual))
{
/*
* CdbLocusType_SegmentGeneral is only used by replicated table
* right now, so if both input and target are replicated table,
* no need to add a motion.
*
* Also, to expand a replicated table to new segments, gpexpand
* force a data reorganization by a query like:
* CREATE TABLE tmp_tab AS SELECT * FROM source_table DISTRIBUTED REPLICATED
* Obviously, tmp_tab in new segments can't get data if we don't
* add a broadcast here.
*/
if(CdbPathLocus_IsSegmentGeneral(subpath->locus) &&
subpath->locus.numsegments >= policy->numsegments)
{
/*
* A query to reach here:
* INSERT INTO d1 SELECT * FROM d1;
* There is no need to add a motion from General, we
* could simply put General on the same segments with
* target table.
*
* Otherwise a broadcast motion is needed otherwise d2 will
* only have data on segment 0.
*/
subpath->locus.numsegments = policy->numsegments;
return subpath;
}
/* plan's data are available on all segment, no motion needed */
if(CdbPathLocus_IsGeneral(subpath->locus))
{
/*
* A query to reach here: INSERT INTO d1 VALUES(1).
* There is no need to add a motion from General, we
* could simply put General on the same segments with
* target table.
*/
subpath->locus.numsegments = Min(subpath->locus.numsegments,policy->numsegments) ;
return subpath;
}
}
subpath = cdbpath_create_broadcast_motion_path(root, subpath, policy->numsegments);
}
else
elog(ERROR, "unrecognized policy type %u", policyType);
return subpath;
}
/*
* Add a suitable Motion Path for delete and update. If the UPDATE
* modifies the distribution key columns, use create_split_update_path()
* instead.
*/
Path *
create_motion_path_for_upddel(PlannerInfo *root, Index rti, GpPolicy *policy,
Path *subpath)
{
GpPolicyType policyType = policy->ptype;
CdbPathLocus targetLocus;
if (policyType == POLICYTYPE_PARTITIONED)
{
if (can_elide_explicit_motion(root, rti, subpath, policy))
return subpath;
else
{
/* GPDB_96_MERGE_FIXME: avoid creating the Explicit Motion in
* simple cases, where all the input data is already on the
* same segment.
*
* Is "strewn" correct here? Can we do better?
*/
CdbPathLocus_MakeStrewn(&targetLocus, policy->numsegments);
subpath = cdbpath_create_explicit_motion_path(root,
subpath,
targetLocus);
}
}
else if (policyType == POLICYTYPE_ENTRY)
{
/* Master-only table */
CdbPathLocus_MakeEntry(&targetLocus);
subpath = cdbpath_create_motion_path(root, subpath, NIL, false, targetLocus);
}
else if (policyType == POLICYTYPE_REPLICATED)
{
/*
* The statement that update or delete on replicated table has to
* be dispatched to each segment and executed on each segment. Thus
* the targetlist cannot contain volatile functions.
*/
if (contain_volatile_functions((Node *) (subpath->pathtarget->exprs)))
elog(ERROR, "could not devise a plan.");
}
else
elog(ERROR, "unrecognized policy type %u", policyType);
return subpath;
}
/*
* In Postgres planner, we add a SplitUpdate node at top so that updating on
* distribution columns could be handled. The SplitUpdate will split each
* update into delete + insert.
*
* There are several important points should be highlighted:
*
* First, in order to split each update operation into two operations,
* delete + insert, we need several junk columns in the subplan's targetlist,
* in addition to the row's new values:
*
* ctid the tuple id used for deletion
*
* gp_segment_id the segment that the row originates from. Usually the
* current segment where the SplitUpdate runs, but not
* necessarily, if there are multiple joins involved and the
* planner decided redistribute the data.
*
* oid if result relation has oids, the old OID, so that it can be
* preserved in the new row.
*
* We will add one more column to the output, the "action". It's an integer
* that indicates for each row, whether it represents the DELETE or the INSERT
* of that row. It is generated by the Split Update node.
*
* Second, current GPDB executor don't support statement-level update triggers
* and will skip row-level update triggers because a split-update is actually
* consist of a delete and insert. So, if the result relation has update
* triggers, we should reject and error out because it's not functional.
*
* GPDB_96_MERGE_FIXME: the below comment is obsolete. Nowadays, SplitUpdate
* computes the new row's hash, and the corresponding. target segment. The
* old segment comes from the gp_segment_id junk column. But ORCA still
* does it the old way!
*
* Third, to support deletion, and hash delete operation to correct segment,
* we need to get attributes of OLD tuple. The old attributes must therefore
* be present in the subplan's target list. That is handled earlier in the
* planner, in expand_targetlist().
*
* For example, a typical plan would be as following for statement:
* update foo set id = l.v + 1 from dep l where foo.v = l.id:
*
* |-- join ( targetlist: [ l.v + 1, foo.v, foo.id, foo.ctid, foo.gp_segment_id ] )
* |
* |-- motion ( targetlist: [l.id, l.v] )
* | |
* | |-- seqscan on dep ....
* |
* |-- hash (targetlist [ v, foo.ctid, foo.gp_segment_id ] )
* |
* |-- seqscan on foo (targetlist: [ v, foo.id, foo.ctid, foo.gp_segment_id ] )
*
* From the plan above, the target foo.id is assigned as l.v + 1, and expand_targetlist()
* ensured that the old value of id, is also available, even though it would not otherwise
* be needed.
*
* 'rti' is the UPDATE target relation.
*/
Path *
create_split_update_path(PlannerInfo *root, Index rti, GpPolicy *policy, Path *subpath)
{
GpPolicyType policyType = policy->ptype;
CdbPathLocus targetLocus;
if (policyType == POLICYTYPE_PARTITIONED)
{
/*
* If any of the distribution key columns are being changed,
* the UPDATE might move tuples from one segment to another.
* Create a Split Update node to deal with that.
*
* If the input is a dummy plan that cannot return any rows,
* e.g. because the input was eliminated by constraint
* exclusion, we can skip it.
*/
targetLocus = cdbpathlocus_for_insert(root, policy, subpath->pathtarget);
subpath = (Path *) make_splitupdate_path(root, subpath, rti);
subpath = cdbpath_create_explicit_motion_path(root,
subpath,
targetLocus);
}
else if (policyType == POLICYTYPE_ENTRY)
{
/* Master-only table */
CdbPathLocus_MakeEntry(&targetLocus);
subpath = cdbpath_create_motion_path(root, subpath, NIL, false, targetLocus);
}
else if (policyType == POLICYTYPE_REPLICATED)
{
}
else
elog(ERROR, "unrecognized policy type %u", policyType);
return subpath;
}
/*
* turn_volatile_seggen_to_singleqe
*
* This function is the key tool to build correct plan
* for general or segmentgeneral locus paths that contain
* volatile functions.
*
* If we find such a pattern:
* 1. if we are update or delete statement on replicated table
* simply reject the query
* 2. if it is general locus, simply change it to singleQE
* 3. if it is segmentgeneral, use a motion to bring it to
* singleQE and then create a projection path
*
* If we do not find the pattern, simply return the input path.
*
* The last parameter of this function is the part that we want to
* check volatile functions.
*/
Path *
turn_volatile_seggen_to_singleqe(PlannerInfo *root, Path *path, Node *node)
{
if ((CdbPathLocus_IsSegmentGeneral(path->locus) || CdbPathLocus_IsGeneral(path->locus)) &&
(contain_volatile_functions(node) || IsA(path, LimitPath)))
{
CdbPathLocus singleQE;
Path *mpath;
ProjectionPath *ppath;
if (root->upd_del_replicated_table > 0 &&
bms_is_member(root->upd_del_replicated_table,
path->parent->relids))
elog(ERROR, "could not devise a plan");
if (CdbPathLocus_IsGeneral(path->locus))
{
CdbPathLocus_MakeSingleQE(&(path->locus),
getgpsegmentCount());
return path;
}
CdbPathLocus_MakeSingleQE(&singleQE,
CdbPathLocus_NumSegments(path->locus));
mpath = cdbpath_create_motion_path(root, path, NIL, false, singleQE);
/*
* mpath might be NULL, like path contain outer Params
* See Github Issue 13532 for details.
*/
if (mpath == NULL)
return path;
ppath = create_projection_path_with_quals(root, mpath->parent, mpath,
mpath->pathtarget, NIL, false);
ppath->force = true;
return (Path *) ppath;
}
else
return path;
}
static SplitUpdatePath *
make_splitupdate_path(PlannerInfo *root, Path *subpath, Index rti)
{
RangeTblEntry *rte;
PathTarget *splitUpdatePathTarget;
SplitUpdatePath *splitupdatepath;
DMLActionExpr *actionExpr;
/* Suppose we already hold locks before caller */
rte = planner_rt_fetch(rti, root);
/*
* Firstly, Trigger is not supported officially by Greenplum.
*
* Secondly, the update trigger is processed in ExecUpdate.
* however, splitupdate will execute ExecSplitUpdate_Insert
* or ExecDelete instead of ExecUpdate. So the update trigger
* will not be triggered in a split plan.
*
* PostgreSQL fires the row-level DELETE, INSERT, and BEFORE
* UPDATE triggers, but not row-level AFTER UPDATE triggers,
* if you UPDATE a partitioning key column.
* Doing a similar thing doesn't help Greenplum likely, the
* behavior would be uncertain since some triggers happen on
* segments and they may require cross segments data changes.
*
* So an update trigger is not allowed when updating the
* distribution key.
*/
if (has_update_triggers(rte->relid))
ereport(ERROR,
(errcode(ERRCODE_GP_FEATURE_NOT_YET),
errmsg("UPDATE on distributed key column not allowed on relation with update triggers")));
/* Add action column at the end of targetlist */
actionExpr = makeNode(DMLActionExpr);
splitUpdatePathTarget = copy_pathtarget(subpath->pathtarget);
add_column_to_pathtarget(splitUpdatePathTarget, (Expr *) actionExpr, 0);
/* populate information generated above into splitupdate node */
splitupdatepath = makeNode(SplitUpdatePath);
splitupdatepath->path.pathtype = T_SplitUpdate;
splitupdatepath->path.parent = subpath->parent;
splitupdatepath->path.pathtarget = splitUpdatePathTarget;
splitupdatepath->path.param_info = NULL;
splitupdatepath->path.parallel_aware = false;
splitupdatepath->path.parallel_safe = subpath->parallel_safe;
splitupdatepath->path.parallel_workers = subpath->parallel_workers;
splitupdatepath->path.rows = 2 * subpath->rows;
splitupdatepath->path.startup_cost = subpath->startup_cost;
splitupdatepath->path.total_cost = subpath->total_cost;
splitupdatepath->path.pathkeys = subpath->pathkeys;
splitupdatepath->path.locus = subpath->locus;
splitupdatepath->subpath = subpath;
splitupdatepath->resultRelation = rti;
return splitupdatepath;
}
static bool
can_elide_explicit_motion(PlannerInfo *root, Index rti, Path *subpath,
GpPolicy *policy)
{
/*
* If there are no Motions between scan of the target relation and here,
* no motion is required.
*/
if (bms_is_member(rti, subpath->sameslice_relids))
return true;
if (!CdbPathLocus_IsStrewn(subpath->locus))
{
CdbPathLocus resultrelation_locus = cdbpathlocus_from_policy(root, rti, policy);
return cdbpathlocus_equal(subpath->locus, resultrelation_locus);
}
return false;
}
相关信息
相关文章
greenplumn cdbappendonlystorageformat 源码
greenplumn cdbappendonlystorageread 源码
greenplumn cdbappendonlystoragewrite 源码
greenplumn cdbappendonlyxlog 源码
greenplumn cdbbufferedappend 源码
0
赞
热门推荐
-
2、 - 优质文章
-
3、 gate.io
-
8、 golang
-
9、 openharmony
-
10、 Vue中input框自动聚焦