1 /*-------------------------------------------------------------------------
4 * foreign-data wrapper for server-side flat files.
6 * Copyright (c) 2010-2015, PostgreSQL Global Development Group
9 * contrib/file_fdw/file_fdw.c
11 *-------------------------------------------------------------------------
18 #include "access/htup_details.h"
19 #include "access/reloptions.h"
20 #include "access/sysattr.h"
21 #include "catalog/pg_foreign_table.h"
22 #include "commands/copy.h"
23 #include "commands/defrem.h"
24 #include "commands/explain.h"
25 #include "commands/vacuum.h"
26 #include "foreign/fdwapi.h"
27 #include "foreign/foreign.h"
28 #include "miscadmin.h"
29 #include "nodes/makefuncs.h"
30 #include "optimizer/cost.h"
31 #include "optimizer/pathnode.h"
32 #include "optimizer/planmain.h"
33 #include "optimizer/restrictinfo.h"
34 #include "optimizer/var.h"
35 #include "utils/memutils.h"
36 #include "utils/rel.h"
37 #include "utils/sampling.h"
42 * Describes the valid options for objects that use this wrapper.
47 Oid optcontext; /* Oid of catalog in which option may appear */
51 * Valid options for file_fdw.
52 * These options are based on the options for the COPY FROM command.
53 * But note that force_not_null and force_null are handled as boolean options
54 * attached to a column, not as table options.
56 * Note: If you are adding new option for user mapping, you need to modify
57 * fileGetOptions(), which currently doesn't bother to look at user mappings.
59 static const struct FileFdwOption valid_options[] = {
61 {"filename", ForeignTableRelationId},
64 /* oids option is not supported */
65 {"format", ForeignTableRelationId},
66 {"header", ForeignTableRelationId},
67 {"delimiter", ForeignTableRelationId},
68 {"quote", ForeignTableRelationId},
69 {"escape", ForeignTableRelationId},
70 {"null", ForeignTableRelationId},
71 {"encoding", ForeignTableRelationId},
72 {"force_not_null", AttributeRelationId},
73 {"force_null", AttributeRelationId},
76 * force_quote is not supported by file_fdw because it's for COPY TO.
84 * FDW-specific information for RelOptInfo.fdw_private.
86 typedef struct FileFdwPlanState
88 char *filename; /* file to read */
89 List *options; /* merged COPY options, excluding filename */
90 BlockNumber pages; /* estimate of file's physical size */
91 double ntuples; /* estimate of number of rows in file */
95 * FDW-specific information for ForeignScanState.fdw_state.
97 typedef struct FileFdwExecutionState
99 char *filename; /* file to read */
100 List *options; /* merged COPY options, excluding filename */
101 CopyState cstate; /* state of reading file */
102 } FileFdwExecutionState;
107 PG_FUNCTION_INFO_V1(file_fdw_handler);
108 PG_FUNCTION_INFO_V1(file_fdw_validator);
111 * FDW callback routines
113 static void fileGetForeignRelSize(PlannerInfo *root,
116 static void fileGetForeignPaths(PlannerInfo *root,
119 static ForeignScan *fileGetForeignPlan(PlannerInfo *root,
122 ForeignPath *best_path,
125 static void fileExplainForeignScan(ForeignScanState *node, ExplainState *es);
126 static void fileBeginForeignScan(ForeignScanState *node, int eflags);
127 static TupleTableSlot *fileIterateForeignScan(ForeignScanState *node);
128 static void fileReScanForeignScan(ForeignScanState *node);
129 static void fileEndForeignScan(ForeignScanState *node);
130 static bool fileAnalyzeForeignTable(Relation relation,
131 AcquireSampleRowsFunc *func,
132 BlockNumber *totalpages);
137 static bool is_valid_option(const char *option, Oid context);
138 static void fileGetOptions(Oid foreigntableid,
139 char **filename, List **other_options);
140 static List *get_file_fdw_attribute_options(Oid relid);
141 static bool check_selective_binary_conversion(RelOptInfo *baserel,
144 static void estimate_size(PlannerInfo *root, RelOptInfo *baserel,
145 FileFdwPlanState *fdw_private);
146 static void estimate_costs(PlannerInfo *root, RelOptInfo *baserel,
147 FileFdwPlanState *fdw_private,
148 Cost *startup_cost, Cost *total_cost);
149 static int file_acquire_sample_rows(Relation onerel, int elevel,
150 HeapTuple *rows, int targrows,
151 double *totalrows, double *totaldeadrows);
155 * Foreign-data wrapper handler function: return a struct with pointers
156 * to my callback routines.
159 file_fdw_handler(PG_FUNCTION_ARGS)
161 FdwRoutine *fdwroutine = makeNode(FdwRoutine);
163 fdwroutine->GetForeignRelSize = fileGetForeignRelSize;
164 fdwroutine->GetForeignPaths = fileGetForeignPaths;
165 fdwroutine->GetForeignPlan = fileGetForeignPlan;
166 fdwroutine->ExplainForeignScan = fileExplainForeignScan;
167 fdwroutine->BeginForeignScan = fileBeginForeignScan;
168 fdwroutine->IterateForeignScan = fileIterateForeignScan;
169 fdwroutine->ReScanForeignScan = fileReScanForeignScan;
170 fdwroutine->EndForeignScan = fileEndForeignScan;
171 fdwroutine->AnalyzeForeignTable = fileAnalyzeForeignTable;
173 PG_RETURN_POINTER(fdwroutine);
177 * Validate the generic options given to a FOREIGN DATA WRAPPER, SERVER,
178 * USER MAPPING or FOREIGN TABLE that uses file_fdw.
180 * Raise an ERROR if the option or its value is considered invalid.
183 file_fdw_validator(PG_FUNCTION_ARGS)
185 List *options_list = untransformRelOptions(PG_GETARG_DATUM(0));
186 Oid catalog = PG_GETARG_OID(1);
187 char *filename = NULL;
188 DefElem *force_not_null = NULL;
189 DefElem *force_null = NULL;
190 List *other_options = NIL;
194 * Only superusers are allowed to set options of a file_fdw foreign table.
195 * This is because the filename is one of those options, and we don't want
196 * non-superusers to be able to determine which file gets read.
198 * Putting this sort of permissions check in a validator is a bit of a
199 * crock, but there doesn't seem to be any other place that can enforce
200 * the check more cleanly.
202 * Note that the valid_options[] array disallows setting filename at any
203 * options level other than foreign table --- otherwise there'd still be a
206 if (catalog == ForeignTableRelationId && !superuser())
208 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
209 errmsg("only superuser can change options of a file_fdw foreign table")));
212 * Check that only options supported by file_fdw, and allowed for the
213 * current object type, are given.
215 foreach(cell, options_list)
217 DefElem *def = (DefElem *) lfirst(cell);
219 if (!is_valid_option(def->defname, catalog))
221 const struct FileFdwOption *opt;
225 * Unknown option specified, complain about it. Provide a hint
226 * with list of valid options for the object.
228 initStringInfo(&buf);
229 for (opt = valid_options; opt->optname; opt++)
231 if (catalog == opt->optcontext)
232 appendStringInfo(&buf, "%s%s", (buf.len > 0) ? ", " : "",
237 (errcode(ERRCODE_FDW_INVALID_OPTION_NAME),
238 errmsg("invalid option \"%s\"", def->defname),
240 ? errhint("Valid options in this context are: %s",
242 : errhint("There are no valid options in this context.")));
246 * Separate out filename and column-specific options, since
247 * ProcessCopyOptions won't accept them.
250 if (strcmp(def->defname, "filename") == 0)
254 (errcode(ERRCODE_SYNTAX_ERROR),
255 errmsg("conflicting or redundant options")));
256 filename = defGetString(def);
260 * force_not_null is a boolean option; after validation we can discard
261 * it - it will be retrieved later in get_file_fdw_attribute_options()
263 else if (strcmp(def->defname, "force_not_null") == 0)
267 (errcode(ERRCODE_SYNTAX_ERROR),
268 errmsg("conflicting or redundant options"),
269 errhint("option \"force_not_null\" supplied more than once for a column")));
270 force_not_null = def;
271 /* Don't care what the value is, as long as it's a legal boolean */
272 (void) defGetBoolean(def);
274 /* See comments for force_not_null above */
275 else if (strcmp(def->defname, "force_null") == 0)
279 (errcode(ERRCODE_SYNTAX_ERROR),
280 errmsg("conflicting or redundant options"),
281 errhint("option \"force_null\" supplied more than once for a column")));
283 (void) defGetBoolean(def);
286 other_options = lappend(other_options, def);
290 * Now apply the core COPY code's validation logic for more checks.
292 ProcessCopyOptions(NULL, true, other_options);
295 * Filename option is required for file_fdw foreign tables.
297 if (catalog == ForeignTableRelationId && filename == NULL)
299 (errcode(ERRCODE_FDW_DYNAMIC_PARAMETER_VALUE_NEEDED),
300 errmsg("filename is required for file_fdw foreign tables")));
306 * Check if the provided option is one of the valid options.
307 * context is the Oid of the catalog holding the object the option is for.
310 is_valid_option(const char *option, Oid context)
312 const struct FileFdwOption *opt;
314 for (opt = valid_options; opt->optname; opt++)
316 if (context == opt->optcontext && strcmp(opt->optname, option) == 0)
323 * Fetch the options for a file_fdw foreign table.
325 * We have to separate out "filename" from the other options because
326 * it must not appear in the options list passed to the core COPY code.
329 fileGetOptions(Oid foreigntableid,
330 char **filename, List **other_options)
333 ForeignServer *server;
334 ForeignDataWrapper *wrapper;
340 * Extract options from FDW objects. We ignore user mappings because
341 * file_fdw doesn't have any options that can be specified there.
343 * (XXX Actually, given the current contents of valid_options[], there's
344 * no point in examining anything except the foreign table's own options.
347 table = GetForeignTable(foreigntableid);
348 server = GetForeignServer(table->serverid);
349 wrapper = GetForeignDataWrapper(server->fdwid);
352 options = list_concat(options, wrapper->options);
353 options = list_concat(options, server->options);
354 options = list_concat(options, table->options);
355 options = list_concat(options, get_file_fdw_attribute_options(foreigntableid));
358 * Separate out the filename.
364 DefElem *def = (DefElem *) lfirst(lc);
366 if (strcmp(def->defname, "filename") == 0)
368 *filename = defGetString(def);
369 options = list_delete_cell(options, lc, prev);
376 * The validator should have checked that a filename was included in the
377 * options, but check again, just in case.
379 if (*filename == NULL)
380 elog(ERROR, "filename is required for file_fdw foreign tables");
382 *other_options = options;
386 * Retrieve per-column generic options from pg_attribute and construct a list
387 * of DefElems representing them.
389 * At the moment we only have "force_not_null", and "force_null",
390 * which should each be combined into a single DefElem listing all such
391 * columns, since that's what COPY expects.
394 get_file_fdw_attribute_options(Oid relid)
400 List *fnncolumns = NIL;
401 List *fncolumns = NIL;
405 rel = heap_open(relid, AccessShareLock);
406 tupleDesc = RelationGetDescr(rel);
407 natts = tupleDesc->natts;
409 /* Retrieve FDW options for all user-defined attributes. */
410 for (attnum = 1; attnum <= natts; attnum++)
412 Form_pg_attribute attr = tupleDesc->attrs[attnum - 1];
416 /* Skip dropped attributes. */
417 if (attr->attisdropped)
420 options = GetForeignColumnOptions(relid, attnum);
423 DefElem *def = (DefElem *) lfirst(lc);
425 if (strcmp(def->defname, "force_not_null") == 0)
427 if (defGetBoolean(def))
429 char *attname = pstrdup(NameStr(attr->attname));
431 fnncolumns = lappend(fnncolumns, makeString(attname));
434 else if (strcmp(def->defname, "force_null") == 0)
436 if (defGetBoolean(def))
438 char *attname = pstrdup(NameStr(attr->attname));
440 fncolumns = lappend(fncolumns, makeString(attname));
443 /* maybe in future handle other options here */
447 heap_close(rel, AccessShareLock);
450 * Return DefElem only when some column(s) have force_not_null /
451 * force_null options set
453 if (fnncolumns != NIL)
454 options = lappend(options, makeDefElem("force_not_null", (Node *) fnncolumns));
456 if (fncolumns != NIL)
457 options = lappend(options, makeDefElem("force_null", (Node *) fncolumns));
463 * fileGetForeignRelSize
464 * Obtain relation size estimates for a foreign table
467 fileGetForeignRelSize(PlannerInfo *root,
471 FileFdwPlanState *fdw_private;
474 * Fetch options. We only need filename at this point, but we might as
475 * well get everything and not need to re-fetch it later in planning.
477 fdw_private = (FileFdwPlanState *) palloc(sizeof(FileFdwPlanState));
478 fileGetOptions(foreigntableid,
479 &fdw_private->filename, &fdw_private->options);
480 baserel->fdw_private = (void *) fdw_private;
482 /* Estimate relation size */
483 estimate_size(root, baserel, fdw_private);
487 * fileGetForeignPaths
488 * Create possible access paths for a scan on the foreign table
490 * Currently we don't support any push-down feature, so there is only one
491 * possible access path, which simply returns all records in the order in
495 fileGetForeignPaths(PlannerInfo *root,
499 FileFdwPlanState *fdw_private = (FileFdwPlanState *) baserel->fdw_private;
503 List *coptions = NIL;
505 /* Decide whether to selectively perform binary conversion */
506 if (check_selective_binary_conversion(baserel,
509 coptions = list_make1(makeDefElem("convert_selectively",
513 estimate_costs(root, baserel, fdw_private,
514 &startup_cost, &total_cost);
517 * Create a ForeignPath node and add it as only possible path. We use the
518 * fdw_private list of the path to carry the convert_selectively option;
519 * it will be propagated into the fdw_private list of the Plan node.
521 add_path(baserel, (Path *)
522 create_foreignscan_path(root, baserel,
526 NIL, /* no pathkeys */
527 NULL, /* no outer rel either */
531 * If data file was sorted, and we knew it somehow, we could insert
532 * appropriate pathkeys into the ForeignPath node to tell the planner
539 * Create a ForeignScan plan node for scanning the foreign table
542 fileGetForeignPlan(PlannerInfo *root,
545 ForeignPath *best_path,
549 Index scan_relid = baserel->relid;
552 * We have no native ability to evaluate restriction clauses, so we just
553 * put all the scan_clauses into the plan node's qual list for the
554 * executor to check. So all we have to do here is strip RestrictInfo
555 * nodes from the clauses and ignore pseudoconstants (which will be
556 * handled elsewhere).
558 scan_clauses = extract_actual_clauses(scan_clauses, false);
560 /* Create the ForeignScan node */
561 return make_foreignscan(tlist,
564 NIL, /* no expressions to evaluate */
565 best_path->fdw_private,
566 NIL /* no custom tlist */ );
570 * fileExplainForeignScan
571 * Produce extra output for EXPLAIN
574 fileExplainForeignScan(ForeignScanState *node, ExplainState *es)
579 /* Fetch options --- we only need filename at this point */
580 fileGetOptions(RelationGetRelid(node->ss.ss_currentRelation),
581 &filename, &options);
583 ExplainPropertyText("Foreign File", filename, es);
585 /* Suppress file size if we're not showing cost details */
588 struct stat stat_buf;
590 if (stat(filename, &stat_buf) == 0)
591 ExplainPropertyLong("Foreign File Size", (long) stat_buf.st_size,
597 * fileBeginForeignScan
598 * Initiate access to the file by creating CopyState
601 fileBeginForeignScan(ForeignScanState *node, int eflags)
603 ForeignScan *plan = (ForeignScan *) node->ss.ps.plan;
607 FileFdwExecutionState *festate;
610 * Do nothing in EXPLAIN (no ANALYZE) case. node->fdw_state stays NULL.
612 if (eflags & EXEC_FLAG_EXPLAIN_ONLY)
615 /* Fetch options of foreign table */
616 fileGetOptions(RelationGetRelid(node->ss.ss_currentRelation),
617 &filename, &options);
619 /* Add any options from the plan (currently only convert_selectively) */
620 options = list_concat(options, plan->fdw_private);
623 * Create CopyState from FDW options. We always acquire all columns, so
624 * as to match the expected ScanTupleSlot signature.
626 cstate = BeginCopyFrom(node->ss.ss_currentRelation,
633 * Save state in node->fdw_state. We must save enough information to call
634 * BeginCopyFrom() again.
636 festate = (FileFdwExecutionState *) palloc(sizeof(FileFdwExecutionState));
637 festate->filename = filename;
638 festate->options = options;
639 festate->cstate = cstate;
641 node->fdw_state = (void *) festate;
645 * fileIterateForeignScan
646 * Read next record from the data file and store it into the
647 * ScanTupleSlot as a virtual tuple
649 static TupleTableSlot *
650 fileIterateForeignScan(ForeignScanState *node)
652 FileFdwExecutionState *festate = (FileFdwExecutionState *) node->fdw_state;
653 TupleTableSlot *slot = node->ss.ss_ScanTupleSlot;
655 ErrorContextCallback errcallback;
657 /* Set up callback to identify error line number. */
658 errcallback.callback = CopyFromErrorCallback;
659 errcallback.arg = (void *) festate->cstate;
660 errcallback.previous = error_context_stack;
661 error_context_stack = &errcallback;
664 * The protocol for loading a virtual tuple into a slot is first
665 * ExecClearTuple, then fill the values/isnull arrays, then
666 * ExecStoreVirtualTuple. If we don't find another row in the file, we
667 * just skip the last step, leaving the slot empty as required.
669 * We can pass ExprContext = NULL because we read all columns from the
670 * file, so no need to evaluate default expressions.
672 * We can also pass tupleOid = NULL because we don't allow oids for
675 ExecClearTuple(slot);
676 found = NextCopyFrom(festate->cstate, NULL,
677 slot->tts_values, slot->tts_isnull,
680 ExecStoreVirtualTuple(slot);
682 /* Remove error callback. */
683 error_context_stack = errcallback.previous;
689 * fileReScanForeignScan
690 * Rescan table, possibly with new parameters
693 fileReScanForeignScan(ForeignScanState *node)
695 FileFdwExecutionState *festate = (FileFdwExecutionState *) node->fdw_state;
697 EndCopyFrom(festate->cstate);
699 festate->cstate = BeginCopyFrom(node->ss.ss_currentRelation,
708 * Finish scanning foreign table and dispose objects used for this scan
711 fileEndForeignScan(ForeignScanState *node)
713 FileFdwExecutionState *festate = (FileFdwExecutionState *) node->fdw_state;
715 /* if festate is NULL, we are in EXPLAIN; nothing to do */
717 EndCopyFrom(festate->cstate);
721 * fileAnalyzeForeignTable
722 * Test whether analyzing this foreign table is supported
725 fileAnalyzeForeignTable(Relation relation,
726 AcquireSampleRowsFunc *func,
727 BlockNumber *totalpages)
731 struct stat stat_buf;
733 /* Fetch options of foreign table */
734 fileGetOptions(RelationGetRelid(relation), &filename, &options);
737 * Get size of the file. (XXX if we fail here, would it be better to just
738 * return false to skip analyzing the table?)
740 if (stat(filename, &stat_buf) < 0)
742 (errcode_for_file_access(),
743 errmsg("could not stat file \"%s\": %m",
747 * Convert size to pages. Must return at least 1 so that we can tell
748 * later on that pg_class.relpages is not default.
750 *totalpages = (stat_buf.st_size + (BLCKSZ - 1)) / BLCKSZ;
754 *func = file_acquire_sample_rows;
760 * check_selective_binary_conversion
762 * Check to see if it's useful to convert only a subset of the file's columns
763 * to binary. If so, construct a list of the column names to be converted,
764 * return that at *columns, and return TRUE. (Note that it's possible to
765 * determine that no columns need be converted, for instance with a COUNT(*)
766 * query. So we can't use returning a NIL list to indicate failure.)
769 check_selective_binary_conversion(RelOptInfo *baserel,
778 Bitmapset *attrs_used = NULL;
779 bool has_wholerow = false;
783 *columns = NIL; /* default result */
786 * Check format of the file. If binary format, this is irrelevant.
788 table = GetForeignTable(foreigntableid);
789 foreach(lc, table->options)
791 DefElem *def = (DefElem *) lfirst(lc);
793 if (strcmp(def->defname, "format") == 0)
795 char *format = defGetString(def);
797 if (strcmp(format, "binary") == 0)
803 /* Collect all the attributes needed for joins or final output. */
804 pull_varattnos((Node *) baserel->reltargetlist, baserel->relid,
807 /* Add all the attributes used by restriction clauses. */
808 foreach(lc, baserel->baserestrictinfo)
810 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
812 pull_varattnos((Node *) rinfo->clause, baserel->relid,
816 /* Convert attribute numbers to column names. */
817 rel = heap_open(foreigntableid, AccessShareLock);
818 tupleDesc = RelationGetDescr(rel);
820 while ((attnum = bms_first_member(attrs_used)) >= 0)
822 /* Adjust for system attributes. */
823 attnum += FirstLowInvalidHeapAttributeNumber;
831 /* Ignore system attributes. */
835 /* Get user attributes. */
838 Form_pg_attribute attr = tupleDesc->attrs[attnum - 1];
839 char *attname = NameStr(attr->attname);
841 /* Skip dropped attributes (probably shouldn't see any here). */
842 if (attr->attisdropped)
844 *columns = lappend(*columns, makeString(pstrdup(attname)));
848 /* Count non-dropped user attributes while we have the tupdesc. */
850 for (i = 0; i < tupleDesc->natts; i++)
852 Form_pg_attribute attr = tupleDesc->attrs[i];
854 if (attr->attisdropped)
859 heap_close(rel, AccessShareLock);
861 /* If there's a whole-row reference, fail: we need all the columns. */
868 /* If all the user attributes are needed, fail. */
869 if (numattrs == list_length(*columns))
879 * Estimate size of a foreign table.
881 * The main result is returned in baserel->rows. We also set
882 * fdw_private->pages and fdw_private->ntuples for later use in the cost
886 estimate_size(PlannerInfo *root, RelOptInfo *baserel,
887 FileFdwPlanState *fdw_private)
889 struct stat stat_buf;
895 * Get size of the file. It might not be there at plan time, though, in
896 * which case we have to use a default estimate.
898 if (stat(fdw_private->filename, &stat_buf) < 0)
899 stat_buf.st_size = 10 * BLCKSZ;
902 * Convert size to pages for use in I/O cost estimate later.
904 pages = (stat_buf.st_size + (BLCKSZ - 1)) / BLCKSZ;
907 fdw_private->pages = pages;
910 * Estimate the number of tuples in the file.
912 if (baserel->pages > 0)
915 * We have # of pages and # of tuples from pg_class (that is, from a
916 * previous ANALYZE), so compute a tuples-per-page estimate and scale
917 * that by the current file size.
921 density = baserel->tuples / (double) baserel->pages;
922 ntuples = clamp_row_est(density * (double) pages);
927 * Otherwise we have to fake it. We back into this estimate using the
928 * planner's idea of the relation width; which is bogus if not all
929 * columns are being read, not to mention that the text representation
930 * of a row probably isn't the same size as its internal
931 * representation. Possibly we could do something better, but the
932 * real answer to anyone who complains is "ANALYZE" ...
936 tuple_width = MAXALIGN(baserel->width) +
937 MAXALIGN(SizeofHeapTupleHeader);
938 ntuples = clamp_row_est((double) stat_buf.st_size /
939 (double) tuple_width);
941 fdw_private->ntuples = ntuples;
944 * Now estimate the number of rows returned by the scan after applying the
945 * baserestrictinfo quals.
948 clauselist_selectivity(root,
949 baserel->baserestrictinfo,
954 nrows = clamp_row_est(nrows);
956 /* Save the output-rows estimate for the planner */
957 baserel->rows = nrows;
961 * Estimate costs of scanning a foreign table.
963 * Results are returned in *startup_cost and *total_cost.
966 estimate_costs(PlannerInfo *root, RelOptInfo *baserel,
967 FileFdwPlanState *fdw_private,
968 Cost *startup_cost, Cost *total_cost)
970 BlockNumber pages = fdw_private->pages;
971 double ntuples = fdw_private->ntuples;
976 * We estimate costs almost the same way as cost_seqscan(), thus assuming
977 * that I/O costs are equivalent to a regular table file of the same size.
978 * However, we take per-tuple CPU costs as 10x of a seqscan, to account
979 * for the cost of parsing records.
981 run_cost += seq_page_cost * pages;
983 *startup_cost = baserel->baserestrictcost.startup;
984 cpu_per_tuple = cpu_tuple_cost * 10 + baserel->baserestrictcost.per_tuple;
985 run_cost += cpu_per_tuple * ntuples;
986 *total_cost = *startup_cost + run_cost;
990 * file_acquire_sample_rows -- acquire a random sample of rows from the table
992 * Selected rows are returned in the caller-allocated array rows[],
993 * which must have at least targrows entries.
994 * The actual number of rows selected is returned as the function result.
995 * We also count the total number of rows in the file and return it into
996 * *totalrows. Note that *totaldeadrows is always set to 0.
998 * Note that the returned list of rows is not always in order by physical
999 * position in the file. Therefore, correlation estimates derived later
1000 * may be meaningless, but it's OK because we don't use the estimates
1001 * currently (the planner only pays attention to correlation for indexscans).
1004 file_acquire_sample_rows(Relation onerel, int elevel,
1005 HeapTuple *rows, int targrows,
1006 double *totalrows, double *totaldeadrows)
1009 double rowstoskip = -1; /* -1 means not set yet */
1010 ReservoirStateData rstate;
1018 ErrorContextCallback errcallback;
1019 MemoryContext oldcontext = CurrentMemoryContext;
1020 MemoryContext tupcontext;
1023 Assert(targrows > 0);
1025 tupDesc = RelationGetDescr(onerel);
1026 values = (Datum *) palloc(tupDesc->natts * sizeof(Datum));
1027 nulls = (bool *) palloc(tupDesc->natts * sizeof(bool));
1029 /* Fetch options of foreign table */
1030 fileGetOptions(RelationGetRelid(onerel), &filename, &options);
1033 * Create CopyState from FDW options.
1035 cstate = BeginCopyFrom(onerel, filename, false, NIL, options);
1038 * Use per-tuple memory context to prevent leak of memory used to read
1039 * rows from the file with Copy routines.
1041 tupcontext = AllocSetContextCreate(CurrentMemoryContext,
1042 "file_fdw temporary context",
1043 ALLOCSET_DEFAULT_MINSIZE,
1044 ALLOCSET_DEFAULT_INITSIZE,
1045 ALLOCSET_DEFAULT_MAXSIZE);
1047 /* Prepare for sampling rows */
1048 reservoir_init_selection_state(&rstate, targrows);
1050 /* Set up callback to identify error line number. */
1051 errcallback.callback = CopyFromErrorCallback;
1052 errcallback.arg = (void *) cstate;
1053 errcallback.previous = error_context_stack;
1054 error_context_stack = &errcallback;
1060 /* Check for user-requested abort or sleep */
1061 vacuum_delay_point();
1063 /* Fetch next row */
1064 MemoryContextReset(tupcontext);
1065 MemoryContextSwitchTo(tupcontext);
1067 found = NextCopyFrom(cstate, NULL, values, nulls, NULL);
1069 MemoryContextSwitchTo(oldcontext);
1075 * The first targrows sample rows are simply copied into the
1076 * reservoir. Then we start replacing tuples in the sample until we
1077 * reach the end of the relation. This algorithm is from Jeff Vitter's
1078 * paper (see more info in commands/analyze.c).
1080 if (numrows < targrows)
1082 rows[numrows++] = heap_form_tuple(tupDesc, values, nulls);
1087 * t in Vitter's paper is the number of records already processed.
1088 * If we need to compute a new S value, we must use the
1089 * not-yet-incremented value of totalrows as t.
1092 rowstoskip = reservoir_get_next_S(&rstate, *totalrows, targrows);
1094 if (rowstoskip <= 0)
1097 * Found a suitable tuple, so save it, replacing one old tuple
1100 int k = (int) (targrows * sampler_random_fract(rstate.randstate));
1102 Assert(k >= 0 && k < targrows);
1103 heap_freetuple(rows[k]);
1104 rows[k] = heap_form_tuple(tupDesc, values, nulls);
1113 /* Remove error callback. */
1114 error_context_stack = errcallback.previous;
1117 MemoryContextDelete(tupcontext);
1119 EndCopyFrom(cstate);
1125 * Emit some interesting relation info
1128 (errmsg("\"%s\": file contains %.0f rows; "
1129 "%d rows in sample",
1130 RelationGetRelationName(onerel),
1131 *totalrows, numrows)));