| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* matview.c |
| 4 |
* materialized view support |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 7 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 8 |
* |
| 9 |
* |
| 10 |
* IDENTIFICATION |
| 11 |
* src/backend/commands/matview.c |
| 12 |
* |
| 13 |
*------------------------------------------------------------------------- |
| 14 |
*/ |
| 15 |
#include "postgres.h" |
| 16 |
|
| 17 |
#include "access/genam.h" |
| 18 |
#include "access/heapam.h" |
| 19 |
#include "access/htup_details.h" |
| 20 |
#include "access/multixact.h" |
| 21 |
#include "access/tableam.h" |
| 22 |
#include "access/xact.h" |
| 23 |
#include "catalog/indexing.h" |
| 24 |
#include "catalog/namespace.h" |
| 25 |
#include "catalog/pg_am.h" |
| 26 |
#include "catalog/pg_opclass.h" |
| 27 |
#include "commands/cluster.h" |
| 28 |
#include "commands/matview.h" |
| 29 |
#include "commands/tablecmds.h" |
| 30 |
#include "commands/tablespace.h" |
| 31 |
#include "executor/executor.h" |
| 32 |
#include "executor/spi.h" |
| 33 |
#include "miscadmin.h" |
| 34 |
#include "pgstat.h" |
| 35 |
#include "rewrite/rewriteHandler.h" |
| 36 |
#include "storage/lmgr.h" |
| 37 |
#include "tcop/tcopprot.h" |
| 38 |
#include "utils/builtins.h" |
| 39 |
#include "utils/lsyscache.h" |
| 40 |
#include "utils/rel.h" |
| 41 |
#include "utils/snapmgr.h" |
| 42 |
#include "utils/syscache.h" |
| 43 |
|
| 44 |
|
| 45 |
typedef struct |
| 46 |
{ |
| 47 |
DestReceiver pub; /* publicly-known function pointers */ |
| 48 |
Oid transientoid; /* OID of new heap into which to store */ |
| 49 |
/* These fields are filled by transientrel_startup: */ |
| 50 |
Relation transientrel; /* relation to write to */ |
| 51 |
CommandId output_cid; /* cmin to insert in output tuples */ |
| 52 |
int ti_options; /* table_tuple_insert performance options */ |
| 53 |
BulkInsertState bistate; /* bulk insert state */ |
| 54 |
} DR_transientrel; |
| 55 |
|
| 56 |
static int matview_maintenance_depth = 0; |
| 57 |
|
| 58 |
static void transientrel_startup(DestReceiver *self, int operation, TupleDesc typeinfo); |
| 59 |
static bool transientrel_receive(TupleTableSlot *slot, DestReceiver *self); |
| 60 |
static void transientrel_shutdown(DestReceiver *self); |
| 61 |
static void transientrel_destroy(DestReceiver *self); |
| 62 |
static uint64 refresh_matview_datafill(DestReceiver *dest, Query *query, |
| 63 |
const char *queryString, bool is_create); |
| 64 |
static char *make_temptable_name_n(char *tempname, int n); |
| 65 |
static void refresh_by_match_merge(Oid matviewOid, Oid tempOid, Oid relowner, |
| 66 |
int save_sec_context); |
| 67 |
static void refresh_by_heap_swap(Oid matviewOid, Oid OIDNewHeap, char relpersistence); |
| 68 |
static bool is_usable_unique_index(Relation indexRel); |
| 69 |
static void OpenMatViewIncrementalMaintenance(void); |
| 70 |
static void CloseMatViewIncrementalMaintenance(void); |
| 71 |
|
| 72 |
/* |
| 73 |
* SetMatViewPopulatedState |
| 74 |
* Mark a materialized view as populated, or not. |
| 75 |
* |
| 76 |
* NOTE: caller must be holding an appropriate lock on the relation. |
| 77 |
*/ |
| 78 |
void |
| 79 |
SetMatViewPopulatedState(Relation relation, bool newstate) |
| 80 |
{ |
| 81 |
Relation pgrel; |
| 82 |
HeapTuple tuple; |
| 83 |
|
| 84 |
Assert(relation->rd_rel->relkind == RELKIND_MATVIEW); |
| 85 |
|
| 86 |
/* |
| 87 |
* Update relation's pg_class entry. Crucial side-effect: other backends |
| 88 |
* (and this one too!) are sent SI message to make them rebuild relcache |
| 89 |
* entries. |
| 90 |
*/ |
| 91 |
pgrel = table_open(RelationRelationId, RowExclusiveLock); |
| 92 |
tuple = SearchSysCacheCopy1(RELOID, |
| 93 |
ObjectIdGetDatum(RelationGetRelid(relation))); |
| 94 |
if (!HeapTupleIsValid(tuple)) |
| 95 |
elog(ERROR, "cache lookup failed for relation %u", |
| 96 |
RelationGetRelid(relation)); |
| 97 |
|
| 98 |
((Form_pg_class) GETSTRUCT(tuple))->relispopulated = newstate; |
| 99 |
|
| 100 |
CatalogTupleUpdate(pgrel, &tuple->t_self, tuple); |
| 101 |
|
| 102 |
heap_freetuple(tuple); |
| 103 |
table_close(pgrel, RowExclusiveLock); |
| 104 |
|
| 105 |
/* |
| 106 |
* Advance command counter to make the updated pg_class row locally |
| 107 |
* visible. |
| 108 |
*/ |
| 109 |
CommandCounterIncrement(); |
| 110 |
} |
| 111 |
|
| 112 |
/* |
| 113 |
* ExecRefreshMatView -- execute a REFRESH MATERIALIZED VIEW command |
| 114 |
* |
| 115 |
* If WITH NO DATA was specified, this is effectively like a TRUNCATE; |
| 116 |
* otherwise it is like a TRUNCATE followed by an INSERT using the SELECT |
| 117 |
* statement associated with the materialized view. The statement node's |
| 118 |
* skipData field shows whether the clause was used. |
| 119 |
*/ |
| 120 |
ObjectAddress |
| 121 |
ExecRefreshMatView(RefreshMatViewStmt *stmt, const char *queryString, |
| 122 |
QueryCompletion *qc) |
| 123 |
{ |
| 124 |
Oid matviewOid; |
| 125 |
LOCKMODE lockmode; |
| 126 |
|
| 127 |
/* Determine strength of lock needed. */ |
| 128 |
lockmode = stmt->concurrent ? ExclusiveLock : AccessExclusiveLock; |
| 129 |
|
| 130 |
/* |
| 131 |
* Get a lock until end of transaction. |
| 132 |
*/ |
| 133 |
matviewOid = RangeVarGetRelidExtended(stmt->relation, |
| 134 |
lockmode, 0, |
| 135 |
RangeVarCallbackMaintainsTable, |
| 136 |
NULL); |
| 137 |
|
| 138 |
return RefreshMatViewByOid(matviewOid, false, stmt->skipData, |
| 139 |
stmt->concurrent, queryString, qc); |
| 140 |
} |
| 141 |
|
| 142 |
/* |
| 143 |
* RefreshMatViewByOid -- refresh materialized view by OID |
| 144 |
* |
| 145 |
* This refreshes the materialized view by creating a new table and swapping |
| 146 |
* the relfilenumbers of the new table and the old materialized view, so the OID |
| 147 |
* of the original materialized view is preserved. Thus we do not lose GRANT |
| 148 |
* nor references to this materialized view. |
| 149 |
* |
| 150 |
* If skipData is true, this is effectively like a TRUNCATE; otherwise it is |
| 151 |
* like a TRUNCATE followed by an INSERT using the SELECT statement associated |
| 152 |
* with the materialized view. |
| 153 |
* |
| 154 |
* Indexes are rebuilt too, via REINDEX. Since we are effectively bulk-loading |
| 155 |
* the new heap, it's better to create the indexes afterwards than to fill them |
| 156 |
* incrementally while we load. |
| 157 |
* |
| 158 |
* The matview's "populated" state is changed based on whether the contents |
| 159 |
* reflect the result set of the materialized view's query. |
| 160 |
* |
| 161 |
* This is also used to populate the materialized view created by CREATE |
| 162 |
* MATERIALIZED VIEW command. |
| 163 |
*/ |
| 164 |
ObjectAddress |
| 165 |
RefreshMatViewByOid(Oid matviewOid, bool is_create, bool skipData, |
| 166 |
bool concurrent, const char *queryString, |
| 167 |
QueryCompletion *qc) |
| 168 |
{ |
| 169 |
Relation matviewRel; |
| 170 |
RewriteRule *rule; |
| 171 |
List *actions; |
| 172 |
Query *dataQuery; |
| 173 |
Oid tableSpace; |
| 174 |
Oid relowner; |
| 175 |
Oid OIDNewHeap; |
| 176 |
uint64 processed = 0; |
| 177 |
char relpersistence; |
| 178 |
Oid save_userid; |
| 179 |
int save_sec_context; |
| 180 |
int save_nestlevel; |
| 181 |
ObjectAddress address; |
| 182 |
|
| 183 |
matviewRel = table_open(matviewOid, NoLock); |
| 184 |
relowner = matviewRel->rd_rel->relowner; |
| 185 |
|
| 186 |
/* |
| 187 |
* Switch to the owner's userid, so that any functions are run as that |
| 188 |
* user. Also lock down security-restricted operations and arrange to |
| 189 |
* make GUC variable changes local to this command. |
| 190 |
*/ |
| 191 |
GetUserIdAndSecContext(&save_userid, &save_sec_context); |
| 192 |
SetUserIdAndSecContext(relowner, |
| 193 |
save_sec_context | SECURITY_RESTRICTED_OPERATION); |
| 194 |
save_nestlevel = NewGUCNestLevel(); |
| 195 |
RestrictSearchPath(); |
| 196 |
|
| 197 |
/* Make sure it is a materialized view. */ |
| 198 |
if (matviewRel->rd_rel->relkind != RELKIND_MATVIEW) |
| 199 |
ereport(ERROR, |
| 200 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 201 |
errmsg("\"%s\" is not a materialized view", |
| 202 |
RelationGetRelationName(matviewRel)))); |
| 203 |
|
| 204 |
/* Check that CONCURRENTLY is not specified if not populated. */ |
| 205 |
if (concurrent && !RelationIsPopulated(matviewRel)) |
| 206 |
ereport(ERROR, |
| 207 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 208 |
errmsg("CONCURRENTLY cannot be used when the materialized view is not populated"))); |
| 209 |
|
| 210 |
/* Check that conflicting options have not been specified. */ |
| 211 |
if (concurrent && skipData) |
| 212 |
ereport(ERROR, |
| 213 |
(errcode(ERRCODE_SYNTAX_ERROR), |
| 214 |
errmsg("%s and %s options cannot be used together", |
| 215 |
"CONCURRENTLY", "WITH NO DATA"))); |
| 216 |
|
| 217 |
/* |
| 218 |
* Check that everything is correct for a refresh. Problems at this point |
| 219 |
* are internal errors, so elog is sufficient. |
| 220 |
*/ |
| 221 |
if (matviewRel->rd_rel->relhasrules == false || |
| 222 |
matviewRel->rd_rules->numLocks < 1) |
| 223 |
elog(ERROR, |
| 224 |
"materialized view \"%s\" is missing rewrite information", |
| 225 |
RelationGetRelationName(matviewRel)); |
| 226 |
|
| 227 |
if (matviewRel->rd_rules->numLocks > 1) |
| 228 |
elog(ERROR, |
| 229 |
"materialized view \"%s\" has too many rules", |
| 230 |
RelationGetRelationName(matviewRel)); |
| 231 |
|
| 232 |
rule = matviewRel->rd_rules->rules[0]; |
| 233 |
if (rule->event != CMD_SELECT || !(rule->isInstead)) |
| 234 |
elog(ERROR, |
| 235 |
"the rule for materialized view \"%s\" is not a SELECT INSTEAD OF rule", |
| 236 |
RelationGetRelationName(matviewRel)); |
| 237 |
|
| 238 |
actions = rule->actions; |
| 239 |
if (list_length(actions) != 1) |
| 240 |
elog(ERROR, |
| 241 |
"the rule for materialized view \"%s\" is not a single action", |
| 242 |
RelationGetRelationName(matviewRel)); |
| 243 |
|
| 244 |
/* |
| 245 |
* Check that there is a unique index with no WHERE clause on one or more |
| 246 |
* columns of the materialized view if CONCURRENTLY is specified. |
| 247 |
*/ |
| 248 |
if (concurrent) |
| 249 |
{ |
| 250 |
List *indexoidlist = RelationGetIndexList(matviewRel); |
| 251 |
ListCell *indexoidscan; |
| 252 |
bool hasUniqueIndex = false; |
| 253 |
|
| 254 |
Assert(!is_create); |
| 255 |
|
| 256 |
foreach(indexoidscan, indexoidlist) |
| 257 |
{ |
| 258 |
Oid indexoid = lfirst_oid(indexoidscan); |
| 259 |
Relation indexRel; |
| 260 |
|
| 261 |
indexRel = index_open(indexoid, AccessShareLock); |
| 262 |
hasUniqueIndex = is_usable_unique_index(indexRel); |
| 263 |
index_close(indexRel, AccessShareLock); |
| 264 |
if (hasUniqueIndex) |
| 265 |
break; |
| 266 |
} |
| 267 |
|
| 268 |
list_free(indexoidlist); |
| 269 |
|
| 270 |
if (!hasUniqueIndex) |
| 271 |
ereport(ERROR, |
| 272 |
(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE), |
| 273 |
errmsg("cannot refresh materialized view \"%s\" concurrently", |
| 274 |
quote_qualified_identifier(get_namespace_name(RelationGetNamespace(matviewRel)), |
| 275 |
RelationGetRelationName(matviewRel))), |
| 276 |
errhint("Create a unique index with no WHERE clause on one or more columns of the materialized view."))); |
| 277 |
} |
| 278 |
|
| 279 |
/* |
| 280 |
* The stored query was rewritten at the time of the MV definition, but |
| 281 |
* has not been scribbled on by the planner. |
| 282 |
*/ |
| 283 |
dataQuery = linitial_node(Query, actions); |
| 284 |
|
| 285 |
/* |
| 286 |
* Check for active uses of the relation in the current transaction, such |
| 287 |
* as open scans. |
| 288 |
* |
| 289 |
* NB: We count on this to protect us against problems with refreshing the |
| 290 |
* data using TABLE_INSERT_FROZEN. |
| 291 |
*/ |
| 292 |
CheckTableNotInUse(matviewRel, |
| 293 |
is_create ? "CREATE MATERIALIZED VIEW" : |
| 294 |
"REFRESH MATERIALIZED VIEW"); |
| 295 |
|
| 296 |
/* |
| 297 |
* Tentatively mark the matview as populated or not (this will roll back |
| 298 |
* if we fail later). |
| 299 |
*/ |
| 300 |
SetMatViewPopulatedState(matviewRel, !skipData); |
| 301 |
|
| 302 |
/* Concurrent refresh builds new data in temp tablespace, and does diff. */ |
| 303 |
if (concurrent) |
| 304 |
{ |
| 305 |
tableSpace = GetDefaultTablespace(RELPERSISTENCE_TEMP, false); |
| 306 |
relpersistence = RELPERSISTENCE_TEMP; |
| 307 |
} |
| 308 |
else |
| 309 |
{ |
| 310 |
tableSpace = matviewRel->rd_rel->reltablespace; |
| 311 |
relpersistence = matviewRel->rd_rel->relpersistence; |
| 312 |
} |
| 313 |
|
| 314 |
/* |
| 315 |
* Create the transient table that will receive the regenerated data. Lock |
| 316 |
* it against access by any other process until commit (by which time it |
| 317 |
* will be gone). |
| 318 |
*/ |
| 319 |
OIDNewHeap = make_new_heap(matviewOid, tableSpace, |
| 320 |
matviewRel->rd_rel->relam, |
| 321 |
relpersistence, ExclusiveLock); |
| 322 |
Assert(CheckRelationOidLockedByMe(OIDNewHeap, AccessExclusiveLock, false)); |
| 323 |
|
| 324 |
/* Generate the data, if wanted. */ |
| 325 |
if (!skipData) |
| 326 |
{ |
| 327 |
DestReceiver *dest; |
| 328 |
|
| 329 |
dest = CreateTransientRelDestReceiver(OIDNewHeap); |
| 330 |
processed = refresh_matview_datafill(dest, dataQuery, queryString, |
| 331 |
is_create); |
| 332 |
} |
| 333 |
|
| 334 |
/* Make the matview match the newly generated data. */ |
| 335 |
if (concurrent) |
| 336 |
{ |
| 337 |
int old_depth = matview_maintenance_depth; |
| 338 |
|
| 339 |
PG_TRY(); |
| 340 |
{ |
| 341 |
refresh_by_match_merge(matviewOid, OIDNewHeap, relowner, |
| 342 |
save_sec_context); |
| 343 |
} |
| 344 |
PG_CATCH(); |
| 345 |
{ |
| 346 |
matview_maintenance_depth = old_depth; |
| 347 |
PG_RE_THROW(); |
| 348 |
} |
| 349 |
PG_END_TRY(); |
| 350 |
Assert(matview_maintenance_depth == old_depth); |
| 351 |
} |
| 352 |
else |
| 353 |
{ |
| 354 |
refresh_by_heap_swap(matviewOid, OIDNewHeap, relpersistence); |
| 355 |
|
| 356 |
/* |
| 357 |
* Inform cumulative stats system about our activity: basically, we |
| 358 |
* truncated the matview and inserted some new data. (The concurrent |
| 359 |
* code path above doesn't need to worry about this because the |
| 360 |
* inserts and deletes it issues get counted by lower-level code.) |
| 361 |
*/ |
| 362 |
pgstat_count_truncate(matviewRel); |
| 363 |
if (!skipData) |
| 364 |
pgstat_count_heap_insert(matviewRel, processed); |
| 365 |
} |
| 366 |
|
| 367 |
table_close(matviewRel, NoLock); |
| 368 |
|
| 369 |
/* Roll back any GUC changes */ |
| 370 |
AtEOXact_GUC(false, save_nestlevel); |
| 371 |
|
| 372 |
/* Restore userid and security context */ |
| 373 |
SetUserIdAndSecContext(save_userid, save_sec_context); |
| 374 |
|
| 375 |
ObjectAddressSet(address, RelationRelationId, matviewOid); |
| 376 |
|
| 377 |
/* |
| 378 |
* Save the rowcount so that pg_stat_statements can track the total number |
| 379 |
* of rows processed by REFRESH MATERIALIZED VIEW command. Note that we |
| 380 |
* still don't display the rowcount in the command completion tag output, |
| 381 |
* i.e., the display_rowcount flag of CMDTAG_REFRESH_MATERIALIZED_VIEW |
| 382 |
* command tag is left false in cmdtaglist.h. Otherwise, the change of |
| 383 |
* completion tag output might break applications using it. |
| 384 |
* |
| 385 |
* When called from CREATE MATERIALIZED VIEW command, the rowcount is |
| 386 |
* displayed with the command tag CMDTAG_SELECT. |
| 387 |
*/ |
| 388 |
if (qc) |
| 389 |
SetQueryCompletion(qc, |
| 390 |
is_create ? CMDTAG_SELECT : CMDTAG_REFRESH_MATERIALIZED_VIEW, |
| 391 |
processed); |
| 392 |
|
| 393 |
return address; |
| 394 |
} |
| 395 |
|
| 396 |
/* |
| 397 |
* refresh_matview_datafill |
| 398 |
* |
| 399 |
* Execute the given query, sending result rows to "dest" (which will |
| 400 |
* insert them into the target matview). |
| 401 |
* |
| 402 |
* Returns number of rows inserted. |
| 403 |
*/ |
| 404 |
static uint64 |
| 405 |
refresh_matview_datafill(DestReceiver *dest, Query *query, |
| 406 |
const char *queryString, bool is_create) |
| 407 |
{ |
| 408 |
List *rewritten; |
| 409 |
PlannedStmt *plan; |
| 410 |
QueryDesc *queryDesc; |
| 411 |
Query *copied_query; |
| 412 |
uint64 processed; |
| 413 |
|
| 414 |
/* Lock and rewrite, using a copy to preserve the original query. */ |
| 415 |
copied_query = copyObject(query); |
| 416 |
AcquireRewriteLocks(copied_query, true, false); |
| 417 |
rewritten = QueryRewrite(copied_query); |
| 418 |
|
| 419 |
/* SELECT should never rewrite to more or less than one SELECT query */ |
| 420 |
if (list_length(rewritten) != 1) |
| 421 |
elog(ERROR, "unexpected rewrite result for %s", |
| 422 |
is_create ? "CREATE MATERIALIZED VIEW " : "REFRESH MATERIALIZED VIEW"); |
| 423 |
query = (Query *) linitial(rewritten); |
| 424 |
|
| 425 |
/* Check for user-requested abort. */ |
| 426 |
CHECK_FOR_INTERRUPTS(); |
| 427 |
|
| 428 |
/* Plan the query which will generate data for the refresh. */ |
| 429 |
plan = pg_plan_query(query, queryString, CURSOR_OPT_PARALLEL_OK, NULL); |
| 430 |
|
| 431 |
/* |
| 432 |
* Use a snapshot with an updated command ID to ensure this query sees |
| 433 |
* results of any previously executed queries. (This could only matter if |
| 434 |
* the planner executed an allegedly-stable function that changed the |
| 435 |
* database contents, but let's do it anyway to be safe.) |
| 436 |
*/ |
| 437 |
PushCopiedSnapshot(GetActiveSnapshot()); |
| 438 |
UpdateActiveSnapshotCommandId(); |
| 439 |
|
| 440 |
/* Create a QueryDesc, redirecting output to our tuple receiver */ |
| 441 |
queryDesc = CreateQueryDesc(plan, queryString, |
| 442 |
GetActiveSnapshot(), InvalidSnapshot, |
| 443 |
dest, NULL, NULL, 0); |
| 444 |
|
| 445 |
/* call ExecutorStart to prepare the plan for execution */ |
| 446 |
ExecutorStart(queryDesc, 0); |
| 447 |
|
| 448 |
/* run the plan */ |
| 449 |
ExecutorRun(queryDesc, ForwardScanDirection, 0); |
| 450 |
|
| 451 |
processed = queryDesc->estate->es_processed; |
| 452 |
|
| 453 |
/* and clean up */ |
| 454 |
ExecutorFinish(queryDesc); |
| 455 |
ExecutorEnd(queryDesc); |
| 456 |
|
| 457 |
FreeQueryDesc(queryDesc); |
| 458 |
|
| 459 |
PopActiveSnapshot(); |
| 460 |
|
| 461 |
return processed; |
| 462 |
} |
| 463 |
|
| 464 |
DestReceiver * |
| 465 |
CreateTransientRelDestReceiver(Oid transientoid) |
| 466 |
{ |
| 467 |
DR_transientrel *self = (DR_transientrel *) palloc0(sizeof(DR_transientrel)); |
| 468 |
|
| 469 |
self->pub.receiveSlot = transientrel_receive; |
| 470 |
self->pub.rStartup = transientrel_startup; |
| 471 |
self->pub.rShutdown = transientrel_shutdown; |
| 472 |
self->pub.rDestroy = transientrel_destroy; |
| 473 |
self->pub.mydest = DestTransientRel; |
| 474 |
self->transientoid = transientoid; |
| 475 |
|
| 476 |
return (DestReceiver *) self; |
| 477 |
} |
| 478 |
|
| 479 |
/* |
| 480 |
* transientrel_startup --- executor startup |
| 481 |
*/ |
| 482 |
static void |
| 483 |
transientrel_startup(DestReceiver *self, int operation, TupleDesc typeinfo) |
| 484 |
{ |
| 485 |
DR_transientrel *myState = (DR_transientrel *) self; |
| 486 |
Relation transientrel; |
| 487 |
|
| 488 |
transientrel = table_open(myState->transientoid, NoLock); |
| 489 |
|
| 490 |
/* |
| 491 |
* Fill private fields of myState for use by later routines |
| 492 |
*/ |
| 493 |
myState->transientrel = transientrel; |
| 494 |
myState->output_cid = GetCurrentCommandId(true); |
| 495 |
myState->ti_options = TABLE_INSERT_SKIP_FSM | TABLE_INSERT_FROZEN; |
| 496 |
myState->bistate = GetBulkInsertState(); |
| 497 |
|
| 498 |
/* |
| 499 |
* Valid smgr_targblock implies something already wrote to the relation. |
| 500 |
* This may be harmless, but this function hasn't planned for it. |
| 501 |
*/ |
| 502 |
Assert(RelationGetTargetBlock(transientrel) == InvalidBlockNumber); |
| 503 |
} |
| 504 |
|
| 505 |
/* |
| 506 |
* transientrel_receive --- receive one tuple |
| 507 |
*/ |
| 508 |
static bool |
| 509 |
transientrel_receive(TupleTableSlot *slot, DestReceiver *self) |
| 510 |
{ |
| 511 |
DR_transientrel *myState = (DR_transientrel *) self; |
| 512 |
|
| 513 |
/* |
| 514 |
* Note that the input slot might not be of the type of the target |
| 515 |
* relation. That's supported by table_tuple_insert(), but slightly less |
| 516 |
* efficient than inserting with the right slot - but the alternative |
| 517 |
* would be to copy into a slot of the right type, which would not be |
| 518 |
* cheap either. This also doesn't allow accessing per-AM data (say a |
| 519 |
* tuple's xmin), but since we don't do that here... |
| 520 |
*/ |
| 521 |
|
| 522 |
table_tuple_insert(myState->transientrel, |
| 523 |
slot, |
| 524 |
myState->output_cid, |
| 525 |
myState->ti_options, |
| 526 |
myState->bistate); |
| 527 |
|
| 528 |
/* We know this is a newly created relation, so there are no indexes */ |
| 529 |
|
| 530 |
return true; |
| 531 |
} |
| 532 |
|
| 533 |
/* |
| 534 |
* transientrel_shutdown --- executor end |
| 535 |
*/ |
| 536 |
static void |
| 537 |
transientrel_shutdown(DestReceiver *self) |
| 538 |
{ |
| 539 |
DR_transientrel *myState = (DR_transientrel *) self; |
| 540 |
|
| 541 |
FreeBulkInsertState(myState->bistate); |
| 542 |
|
| 543 |
table_finish_bulk_insert(myState->transientrel, myState->ti_options); |
| 544 |
|
| 545 |
/* close transientrel, but keep lock until commit */ |
| 546 |
table_close(myState->transientrel, NoLock); |
| 547 |
myState->transientrel = NULL; |
| 548 |
} |
| 549 |
|
| 550 |
/* |
| 551 |
* transientrel_destroy --- release DestReceiver object |
| 552 |
*/ |
| 553 |
static void |
| 554 |
transientrel_destroy(DestReceiver *self) |
| 555 |
{ |
| 556 |
pfree(self); |
| 557 |
} |
| 558 |
|
| 559 |
|
| 560 |
/* |
| 561 |
* Given a qualified temporary table name, append an underscore followed by |
| 562 |
* the given integer, to make a new table name based on the old one. |
| 563 |
* The result is a palloc'd string. |
| 564 |
* |
| 565 |
* As coded, this would fail to make a valid SQL name if the given name were, |
| 566 |
* say, "FOO"."BAR". Currently, the table name portion of the input will |
| 567 |
* never be double-quoted because it's of the form "pg_temp_NNN", cf |
| 568 |
* make_new_heap(). But we might have to work harder someday. |
| 569 |
*/ |
| 570 |
static char * |
| 571 |
make_temptable_name_n(char *tempname, int n) |
| 572 |
{ |
| 573 |
StringInfoData namebuf; |
| 574 |
|
| 575 |
initStringInfo(&namebuf); |
| 576 |
appendStringInfoString(&namebuf, tempname); |
| 577 |
appendStringInfo(&namebuf, "_%d", n); |
| 578 |
return namebuf.data; |
| 579 |
} |
| 580 |
|
| 581 |
/* |
| 582 |
* refresh_by_match_merge |
| 583 |
* |
| 584 |
* Refresh a materialized view with transactional semantics, while allowing |
| 585 |
* concurrent reads. |
| 586 |
* |
| 587 |
* This is called after a new version of the data has been created in a |
| 588 |
* temporary table. It performs a full outer join against the old version of |
| 589 |
* the data, producing "diff" results. This join cannot work if there are any |
| 590 |
* duplicated rows in either the old or new versions, in the sense that every |
| 591 |
* column would compare as equal between the two rows. It does work correctly |
| 592 |
* in the face of rows which have at least one NULL value, with all non-NULL |
| 593 |
* columns equal. The behavior of NULLs on equality tests and on UNIQUE |
| 594 |
* indexes turns out to be quite convenient here; the tests we need to make |
| 595 |
* are consistent with default behavior. If there is at least one UNIQUE |
| 596 |
* index on the materialized view, we have exactly the guarantee we need. |
| 597 |
* |
| 598 |
* The temporary table used to hold the diff results contains just the TID of |
| 599 |
* the old record (if matched) and the ROW from the new table as a single |
| 600 |
* column of complex record type (if matched). |
| 601 |
* |
| 602 |
* Once we have the diff table, we perform set-based DELETE and INSERT |
| 603 |
* operations against the materialized view, and discard both temporary |
| 604 |
* tables. |
| 605 |
* |
| 606 |
* Everything from the generation of the new data to applying the differences |
| 607 |
* takes place under cover of an ExclusiveLock, since it seems as though we |
| 608 |
* would want to prohibit not only concurrent REFRESH operations, but also |
| 609 |
* incremental maintenance. It also doesn't seem reasonable or safe to allow |
| 610 |
* SELECT FOR UPDATE or SELECT FOR SHARE on rows being updated or deleted by |
| 611 |
* this command. |
| 612 |
*/ |
| 613 |
static void |
| 614 |
refresh_by_match_merge(Oid matviewOid, Oid tempOid, Oid relowner, |
| 615 |
int save_sec_context) |
| 616 |
{ |
| 617 |
StringInfoData querybuf; |
| 618 |
Relation matviewRel; |
| 619 |
Relation tempRel; |
| 620 |
char *matviewname; |
| 621 |
char *tempname; |
| 622 |
char *diffname; |
| 623 |
TupleDesc tupdesc; |
| 624 |
bool foundUniqueIndex; |
| 625 |
List *indexoidlist; |
| 626 |
ListCell *indexoidscan; |
| 627 |
int16 relnatts; |
| 628 |
Oid *opUsedForQual; |
| 629 |
|
| 630 |
initStringInfo(&querybuf); |
| 631 |
matviewRel = table_open(matviewOid, NoLock); |
| 632 |
matviewname = quote_qualified_identifier(get_namespace_name(RelationGetNamespace(matviewRel)), |
| 633 |
RelationGetRelationName(matviewRel)); |
| 634 |
tempRel = table_open(tempOid, NoLock); |
| 635 |
tempname = quote_qualified_identifier(get_namespace_name(RelationGetNamespace(tempRel)), |
| 636 |
RelationGetRelationName(tempRel)); |
| 637 |
diffname = make_temptable_name_n(tempname, 2); |
| 638 |
|
| 639 |
relnatts = RelationGetNumberOfAttributes(matviewRel); |
| 640 |
|
| 641 |
/* Open SPI context. */ |
| 642 |
SPI_connect(); |
| 643 |
|
| 644 |
/* Analyze the temp table with the new contents. */ |
| 645 |
appendStringInfo(&querybuf, "ANALYZE %s", tempname); |
| 646 |
if (SPI_exec(querybuf.data, 0) != SPI_OK_UTILITY) |
| 647 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 648 |
|
| 649 |
/* |
| 650 |
* We need to ensure that there are not duplicate rows without NULLs in |
| 651 |
* the new data set before we can count on the "diff" results. Check for |
| 652 |
* that in a way that allows showing the first duplicated row found. Even |
| 653 |
* after we pass this test, a unique index on the materialized view may |
| 654 |
* find a duplicate key problem. |
| 655 |
* |
| 656 |
* Note: here and below, we use "tablename.*::tablerowtype" as a hack to |
| 657 |
* keep ".*" from being expanded into multiple columns in a SELECT list. |
| 658 |
* Compare ruleutils.c's get_variable(). |
| 659 |
*/ |
| 660 |
resetStringInfo(&querybuf); |
| 661 |
appendStringInfo(&querybuf, |
| 662 |
"SELECT newdata.*::%s FROM %s newdata " |
| 663 |
"WHERE newdata.* IS NOT NULL AND EXISTS " |
| 664 |
"(SELECT 1 FROM %s newdata2 WHERE newdata2.* IS NOT NULL " |
| 665 |
"AND newdata2.* OPERATOR(pg_catalog.*=) newdata.* " |
| 666 |
"AND newdata2.ctid OPERATOR(pg_catalog.<>) " |
| 667 |
"newdata.ctid)", |
| 668 |
tempname, tempname, tempname); |
| 669 |
if (SPI_execute(querybuf.data, false, 1) != SPI_OK_SELECT) |
| 670 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 671 |
if (SPI_processed > 0) |
| 672 |
{ |
| 673 |
/* |
| 674 |
* Note that this ereport() is returning data to the user. Generally, |
| 675 |
* we would want to make sure that the user has been granted access to |
| 676 |
* this data. However, REFRESH MAT VIEW is only able to be run by the |
| 677 |
* owner of the mat view (or a superuser) and therefore there is no |
| 678 |
* need to check for access to data in the mat view. |
| 679 |
*/ |
| 680 |
ereport(ERROR, |
| 681 |
(errcode(ERRCODE_CARDINALITY_VIOLATION), |
| 682 |
errmsg("new data for materialized view \"%s\" contains duplicate rows without any null columns", |
| 683 |
RelationGetRelationName(matviewRel)), |
| 684 |
errdetail("Row: %s", |
| 685 |
SPI_getvalue(SPI_tuptable->vals[0], SPI_tuptable->tupdesc, 1)))); |
| 686 |
} |
| 687 |
|
| 688 |
/* |
| 689 |
* Create the temporary "diff" table. |
| 690 |
* |
| 691 |
* Temporarily switch out of the SECURITY_RESTRICTED_OPERATION context, |
| 692 |
* because you cannot create temp tables in SRO context. For extra |
| 693 |
* paranoia, add the composite type column only after switching back to |
| 694 |
* SRO context. |
| 695 |
*/ |
| 696 |
SetUserIdAndSecContext(relowner, |
| 697 |
save_sec_context | SECURITY_LOCAL_USERID_CHANGE); |
| 698 |
resetStringInfo(&querybuf); |
| 699 |
appendStringInfo(&querybuf, |
| 700 |
"CREATE TEMP TABLE %s (tid pg_catalog.tid)", |
| 701 |
diffname); |
| 702 |
if (SPI_exec(querybuf.data, 0) != SPI_OK_UTILITY) |
| 703 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 704 |
SetUserIdAndSecContext(relowner, |
| 705 |
save_sec_context | SECURITY_RESTRICTED_OPERATION); |
| 706 |
resetStringInfo(&querybuf); |
| 707 |
appendStringInfo(&querybuf, |
| 708 |
"ALTER TABLE %s ADD COLUMN newdata %s", |
| 709 |
diffname, tempname); |
| 710 |
if (SPI_exec(querybuf.data, 0) != SPI_OK_UTILITY) |
| 711 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 712 |
|
| 713 |
/* Start building the query for populating the diff table. */ |
| 714 |
resetStringInfo(&querybuf); |
| 715 |
appendStringInfo(&querybuf, |
| 716 |
"INSERT INTO %s " |
| 717 |
"SELECT mv.ctid AS tid, newdata.*::%s AS newdata " |
| 718 |
"FROM %s mv FULL JOIN %s newdata ON (", |
| 719 |
diffname, tempname, matviewname, tempname); |
| 720 |
|
| 721 |
/* |
| 722 |
* Get the list of index OIDs for the table from the relcache, and look up |
| 723 |
* each one in the pg_index syscache. We will test for equality on all |
| 724 |
* columns present in all unique indexes which only reference columns and |
| 725 |
* include all rows. |
| 726 |
*/ |
| 727 |
tupdesc = matviewRel->rd_att; |
| 728 |
opUsedForQual = (Oid *) palloc0(sizeof(Oid) * relnatts); |
| 729 |
foundUniqueIndex = false; |
| 730 |
|
| 731 |
indexoidlist = RelationGetIndexList(matviewRel); |
| 732 |
|
| 733 |
foreach(indexoidscan, indexoidlist) |
| 734 |
{ |
| 735 |
Oid indexoid = lfirst_oid(indexoidscan); |
| 736 |
Relation indexRel; |
| 737 |
|
| 738 |
indexRel = index_open(indexoid, RowExclusiveLock); |
| 739 |
if (is_usable_unique_index(indexRel)) |
| 740 |
{ |
| 741 |
Form_pg_index indexStruct = indexRel->rd_index; |
| 742 |
int indnkeyatts = indexStruct->indnkeyatts; |
| 743 |
oidvector *indclass; |
| 744 |
Datum indclassDatum; |
| 745 |
int i; |
| 746 |
|
| 747 |
/* Must get indclass the hard way. */ |
| 748 |
indclassDatum = SysCacheGetAttrNotNull(INDEXRELID, |
| 749 |
indexRel->rd_indextuple, |
| 750 |
Anum_pg_index_indclass); |
| 751 |
indclass = (oidvector *) DatumGetPointer(indclassDatum); |
| 752 |
|
| 753 |
/* Add quals for all columns from this index. */ |
| 754 |
for (i = 0; i < indnkeyatts; i++) |
| 755 |
{ |
| 756 |
int attnum = indexStruct->indkey.values[i]; |
| 757 |
Oid opclass = indclass->values[i]; |
| 758 |
Form_pg_attribute attr = TupleDescAttr(tupdesc, attnum - 1); |
| 759 |
Oid attrtype = attr->atttypid; |
| 760 |
HeapTuple cla_ht; |
| 761 |
Form_pg_opclass cla_tup; |
| 762 |
Oid opfamily; |
| 763 |
Oid opcintype; |
| 764 |
Oid op; |
| 765 |
const char *leftop; |
| 766 |
const char *rightop; |
| 767 |
|
| 768 |
/* |
| 769 |
* Identify the equality operator associated with this index |
| 770 |
* column. First we need to look up the column's opclass. |
| 771 |
*/ |
| 772 |
cla_ht = SearchSysCache1(CLAOID, ObjectIdGetDatum(opclass)); |
| 773 |
if (!HeapTupleIsValid(cla_ht)) |
| 774 |
elog(ERROR, "cache lookup failed for opclass %u", opclass); |
| 775 |
cla_tup = (Form_pg_opclass) GETSTRUCT(cla_ht); |
| 776 |
opfamily = cla_tup->opcfamily; |
| 777 |
opcintype = cla_tup->opcintype; |
| 778 |
ReleaseSysCache(cla_ht); |
| 779 |
|
| 780 |
op = get_opfamily_member_for_cmptype(opfamily, opcintype, opcintype, COMPARE_EQ); |
| 781 |
if (!OidIsValid(op)) |
| 782 |
elog(ERROR, "missing equality operator for (%u,%u) in opfamily %u", |
| 783 |
opcintype, opcintype, opfamily); |
| 784 |
|
| 785 |
/* |
| 786 |
* If we find the same column with the same equality semantics |
| 787 |
* in more than one index, we only need to emit the equality |
| 788 |
* clause once. |
| 789 |
* |
| 790 |
* Since we only remember the last equality operator, this |
| 791 |
* code could be fooled into emitting duplicate clauses given |
| 792 |
* multiple indexes with several different opclasses ... but |
| 793 |
* that's so unlikely it doesn't seem worth spending extra |
| 794 |
* code to avoid. |
| 795 |
*/ |
| 796 |
if (opUsedForQual[attnum - 1] == op) |
| 797 |
continue; |
| 798 |
opUsedForQual[attnum - 1] = op; |
| 799 |
|
| 800 |
/* |
| 801 |
* Actually add the qual, ANDed with any others. |
| 802 |
*/ |
| 803 |
if (foundUniqueIndex) |
| 804 |
appendStringInfoString(&querybuf, " AND "); |
| 805 |
|
| 806 |
leftop = quote_qualified_identifier("newdata", |
| 807 |
NameStr(attr->attname)); |
| 808 |
rightop = quote_qualified_identifier("mv", |
| 809 |
NameStr(attr->attname)); |
| 810 |
|
| 811 |
generate_operator_clause(&querybuf, |
| 812 |
leftop, attrtype, |
| 813 |
op, |
| 814 |
rightop, attrtype); |
| 815 |
|
| 816 |
foundUniqueIndex = true; |
| 817 |
} |
| 818 |
} |
| 819 |
|
| 820 |
/* Keep the locks, since we're about to run DML which needs them. */ |
| 821 |
index_close(indexRel, NoLock); |
| 822 |
} |
| 823 |
|
| 824 |
list_free(indexoidlist); |
| 825 |
|
| 826 |
/* |
| 827 |
* There must be at least one usable unique index on the matview. |
| 828 |
* |
| 829 |
* ExecRefreshMatView() checks that after taking the exclusive lock on the |
| 830 |
* matview. So at least one unique index is guaranteed to exist here |
| 831 |
* because the lock is still being held. (One known exception is if a |
| 832 |
* function called as part of refreshing the matview drops the index. |
| 833 |
* That's a pretty silly thing to do.) |
| 834 |
*/ |
| 835 |
if (!foundUniqueIndex) |
| 836 |
ereport(ERROR, |
| 837 |
errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 838 |
errmsg("could not find suitable unique index on materialized view \"%s\"", |
| 839 |
RelationGetRelationName(matviewRel))); |
| 840 |
|
| 841 |
appendStringInfoString(&querybuf, |
| 842 |
" AND newdata.* OPERATOR(pg_catalog.*=) mv.*) " |
| 843 |
"WHERE newdata.* IS NULL OR mv.* IS NULL " |
| 844 |
"ORDER BY tid"); |
| 845 |
|
| 846 |
/* Populate the temporary "diff" table. */ |
| 847 |
if (SPI_exec(querybuf.data, 0) != SPI_OK_INSERT) |
| 848 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 849 |
|
| 850 |
/* |
| 851 |
* We have no further use for data from the "full-data" temp table, but we |
| 852 |
* must keep it around because its type is referenced from the diff table. |
| 853 |
*/ |
| 854 |
|
| 855 |
/* Analyze the diff table. */ |
| 856 |
resetStringInfo(&querybuf); |
| 857 |
appendStringInfo(&querybuf, "ANALYZE %s", diffname); |
| 858 |
if (SPI_exec(querybuf.data, 0) != SPI_OK_UTILITY) |
| 859 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 860 |
|
| 861 |
OpenMatViewIncrementalMaintenance(); |
| 862 |
|
| 863 |
/* Deletes must come before inserts; do them first. */ |
| 864 |
resetStringInfo(&querybuf); |
| 865 |
appendStringInfo(&querybuf, |
| 866 |
"DELETE FROM %s mv WHERE ctid OPERATOR(pg_catalog.=) ANY " |
| 867 |
"(SELECT diff.tid FROM %s diff " |
| 868 |
"WHERE diff.tid IS NOT NULL " |
| 869 |
"AND diff.newdata IS NULL)", |
| 870 |
matviewname, diffname); |
| 871 |
if (SPI_exec(querybuf.data, 0) != SPI_OK_DELETE) |
| 872 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 873 |
|
| 874 |
/* Inserts go last. */ |
| 875 |
resetStringInfo(&querybuf); |
| 876 |
appendStringInfo(&querybuf, |
| 877 |
"INSERT INTO %s SELECT (diff.newdata).* " |
| 878 |
"FROM %s diff WHERE tid IS NULL", |
| 879 |
matviewname, diffname); |
| 880 |
if (SPI_exec(querybuf.data, 0) != SPI_OK_INSERT) |
| 881 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 882 |
|
| 883 |
/* We're done maintaining the materialized view. */ |
| 884 |
CloseMatViewIncrementalMaintenance(); |
| 885 |
table_close(tempRel, NoLock); |
| 886 |
table_close(matviewRel, NoLock); |
| 887 |
|
| 888 |
/* Clean up temp tables. */ |
| 889 |
resetStringInfo(&querybuf); |
| 890 |
appendStringInfo(&querybuf, "DROP TABLE %s, %s", diffname, tempname); |
| 891 |
if (SPI_exec(querybuf.data, 0) != SPI_OK_UTILITY) |
| 892 |
elog(ERROR, "SPI_exec failed: %s", querybuf.data); |
| 893 |
|
| 894 |
/* Close SPI context. */ |
| 895 |
if (SPI_finish() != SPI_OK_FINISH) |
| 896 |
elog(ERROR, "SPI_finish failed"); |
| 897 |
} |
| 898 |
|
| 899 |
/* |
| 900 |
* Swap the physical files of the target and transient tables, then rebuild |
| 901 |
* the target's indexes and throw away the transient table. Security context |
| 902 |
* swapping is handled by the called function, so it is not needed here. |
| 903 |
*/ |
| 904 |
static void |
| 905 |
refresh_by_heap_swap(Oid matviewOid, Oid OIDNewHeap, char relpersistence) |
| 906 |
{ |
| 907 |
finish_heap_swap(matviewOid, OIDNewHeap, false, false, true, true, |
| 908 |
RecentXmin, ReadNextMultiXactId(), relpersistence); |
| 909 |
} |
| 910 |
|
| 911 |
/* |
| 912 |
* Check whether specified index is usable for match merge. |
| 913 |
*/ |
| 914 |
static bool |
| 915 |
is_usable_unique_index(Relation indexRel) |
| 916 |
{ |
| 917 |
Form_pg_index indexStruct = indexRel->rd_index; |
| 918 |
|
| 919 |
/* |
| 920 |
* Must be unique, valid, immediate, non-partial, and be defined over |
| 921 |
* plain user columns (not expressions). |
| 922 |
*/ |
| 923 |
if (indexStruct->indisunique && |
| 924 |
indexStruct->indimmediate && |
| 925 |
indexStruct->indisvalid && |
| 926 |
RelationGetIndexPredicate(indexRel) == NIL && |
| 927 |
indexStruct->indnatts > 0) |
| 928 |
{ |
| 929 |
/* |
| 930 |
* The point of groveling through the index columns individually is to |
| 931 |
* reject both index expressions and system columns. Currently, |
| 932 |
* matviews couldn't have OID columns so there's no way to create an |
| 933 |
* index on a system column; but maybe someday that wouldn't be true, |
| 934 |
* so let's be safe. |
| 935 |
*/ |
| 936 |
int numatts = indexStruct->indnatts; |
| 937 |
int i; |
| 938 |
|
| 939 |
for (i = 0; i < numatts; i++) |
| 940 |
{ |
| 941 |
int attnum = indexStruct->indkey.values[i]; |
| 942 |
|
| 943 |
if (attnum <= 0) |
| 944 |
return false; |
| 945 |
} |
| 946 |
return true; |
| 947 |
} |
| 948 |
return false; |
| 949 |
} |
| 950 |
|
| 951 |
|
| 952 |
/* |
| 953 |
* This should be used to test whether the backend is in a context where it is |
| 954 |
* OK to allow DML statements to modify materialized views. We only want to |
| 955 |
* allow that for internal code driven by the materialized view definition, |
| 956 |
* not for arbitrary user-supplied code. |
| 957 |
* |
| 958 |
* While the function names reflect the fact that their main intended use is |
| 959 |
* incremental maintenance of materialized views (in response to changes to |
| 960 |
* the data in referenced relations), they are initially used to allow REFRESH |
| 961 |
* without blocking concurrent reads. |
| 962 |
*/ |
| 963 |
bool |
| 964 |
MatViewIncrementalMaintenanceIsEnabled(void) |
| 965 |
{ |
| 966 |
return matview_maintenance_depth > 0; |
| 967 |
} |
| 968 |
|
| 969 |
static void |
| 970 |
OpenMatViewIncrementalMaintenance(void) |
| 971 |
{ |
| 972 |
matview_maintenance_depth++; |
| 973 |
} |
| 974 |
|
| 975 |
static void |
| 976 |
CloseMatViewIncrementalMaintenance(void) |
| 977 |
{ |
| 978 |
matview_maintenance_depth--; |
| 979 |
Assert(matview_maintenance_depth >= 0); |
| 980 |
} |