database/postgres
1
0
Fork 0
mirror of https://github.com/postgres/postgres.git synced 2025-07-27 12:41:57 +03:00
Code Activity

Revise FDW planning API, again.

Browse Source 
Further reflection shows that a single callback isn't very workable if we
desire to let FDWs generate multiple Paths, because that forces the FDW to
do all work necessary to generate a valid Plan node for each Path.  Instead
split the former PlanForeignScan API into three steps: GetForeignRelSize,
GetForeignPaths, GetForeignPlan.  We had already bit the bullet of breaking
the 9.1 FDW API for 9.2, so this shouldn't cause very much additional pain,
and it's substantially more flexible for complex FDWs.

Add an fdw_private field to RelOptInfo so that the new functions can save
state there rather than possibly having to recalculate information two or
three times.

In addition, we'd not thought through what would be needed to allow an FDW
to set up subexpressions of its choice for runtime execution.  We could
treat ForeignScan.fdw_private as an executable expression but that seems
likely to break existing FDWs unnecessarily (in particular, it would
restrict the set of node types allowable in fdw_private to those supported
by expression_tree_walker).  Instead, invent a separate field fdw_exprs
which will receive the postprocessing appropriate for expression trees.
(One field is enough since it can be a list of expressions; also, we assume
the corresponding expression state tree(s) will be held within fdw_state,
so we don't need to add anything to ForeignScanState.)

Per review of Hanada Shigeru's pgsql_fdw patch.  We may need to tweak this
further as we continue to work on that patch, but to me it feels a lot
closer to being right now.
This commit is contained in:
Tom Lane
2012-03-09 12:48:48 -05:00
parent 342baf4ce6
commit b14953932d
15 changed files with 435 additions and 111 deletions
Download Patch File Download Diff File Expand all files Collapse all files

161
contrib/file_fdw/file_fdw.c
View File

@ -26,6 +26,8 @@
#include "nodes/makefuncs.h"
#include "optimizer/cost.h"
#include "optimizer/pathnode.h"
#include "optimizer/planmain.h"
#include "optimizer/restrictinfo.h"
#include "utils/rel.h"
PG_MODULE_MAGIC;
@ -48,7 +50,7 @@ struct FileFdwOption
* Note: If you are adding new option for user mapping, you need to modify
* fileGetOptions(), which currently doesn't bother to look at user mappings.
*/
static struct FileFdwOption valid_options[] = {
static const struct FileFdwOption valid_options[] = {
/* File options */
{"filename", ForeignTableRelationId},
@ -71,6 +73,17 @@ static struct FileFdwOption valid_options[] = {
{NULL, InvalidOid}
};
/*
* FDW-specific information for RelOptInfo.fdw_private.
*/
typedef struct FileFdwPlanState
{
char *filename; /* file to read */
List *options; /* merged COPY options, excluding filename */
BlockNumber pages; /* estimate of file's physical size */
double ntuples; /* estimate of number of rows in file */
} FileFdwPlanState;
/*
* FDW-specific information for ForeignScanState.fdw_state.
*/
@ -93,9 +106,18 @@ PG_FUNCTION_INFO_V1(file_fdw_validator);
/*
* FDW callback routines
*/
static void filePlanForeignScan(Oid foreigntableid,
PlannerInfo *root,
RelOptInfo *baserel);
static void fileGetForeignRelSize(PlannerInfo *root,
RelOptInfo *baserel,
Oid foreigntableid);
static void fileGetForeignPaths(PlannerInfo *root,
RelOptInfo *baserel,
Oid foreigntableid);
static ForeignScan *fileGetForeignPlan(PlannerInfo *root,
RelOptInfo *baserel,
Oid foreigntableid,
ForeignPath *best_path,
List *tlist,
List *scan_clauses);
static void fileExplainForeignScan(ForeignScanState *node, ExplainState *es);
static void fileBeginForeignScan(ForeignScanState *node, int eflags);
static TupleTableSlot *fileIterateForeignScan(ForeignScanState *node);
@ -109,8 +131,10 @@ static bool is_valid_option(const char *option, Oid context);
static void fileGetOptions(Oid foreigntableid,
char **filename, List **other_options);
static List *get_file_fdw_attribute_options(Oid relid);
static void estimate_size(PlannerInfo *root, RelOptInfo *baserel,
FileFdwPlanState *fdw_private);
static void estimate_costs(PlannerInfo *root, RelOptInfo *baserel,
const char *filename,
FileFdwPlanState *fdw_private,
Cost *startup_cost, Cost *total_cost);
@ -123,7 +147,9 @@ file_fdw_handler(PG_FUNCTION_ARGS)
{
FdwRoutine *fdwroutine = makeNode(FdwRoutine);
fdwroutine->PlanForeignScan = filePlanForeignScan;
fdwroutine->GetForeignRelSize = fileGetForeignRelSize;
fdwroutine->GetForeignPaths = fileGetForeignPaths;
fdwroutine->GetForeignPlan = fileGetForeignPlan;
fdwroutine->ExplainForeignScan = fileExplainForeignScan;
fdwroutine->BeginForeignScan = fileBeginForeignScan;
fdwroutine->IterateForeignScan = fileIterateForeignScan;
@ -177,7 +203,7 @@ file_fdw_validator(PG_FUNCTION_ARGS)
if (!is_valid_option(def->defname, catalog))
{
struct FileFdwOption *opt;
const struct FileFdwOption *opt;
StringInfoData buf;
/*
@ -249,7 +275,7 @@ file_fdw_validator(PG_FUNCTION_ARGS)
static bool
is_valid_option(const char *option, Oid context)
{
struct FileFdwOption *opt;
const struct FileFdwOption *opt;
for (opt = valid_options; opt->optname; opt++)
{
@ -381,7 +407,31 @@ get_file_fdw_attribute_options(Oid relid)
}
/*
* filePlanForeignScan
* fileGetForeignRelSize
* Obtain relation size estimates for a foreign table
*/
static void
fileGetForeignRelSize(PlannerInfo *root,
RelOptInfo *baserel,
Oid foreigntableid)
{
FileFdwPlanState *fdw_private;
/*
* Fetch options. We only need filename at this point, but we might
* as well get everything and not need to re-fetch it later in planning.
*/
fdw_private = (FileFdwPlanState *) palloc(sizeof(FileFdwPlanState));
fileGetOptions(foreigntableid,
&fdw_private->filename, &fdw_private->options);
baserel->fdw_private = (void *) fdw_private;
/* Estimate relation size */
estimate_size(root, baserel, fdw_private);
}
/*
* fileGetForeignPaths
* Create possible access paths for a scan on the foreign table
*
* Currently we don't support any push-down feature, so there is only one
@ -389,20 +439,16 @@ get_file_fdw_attribute_options(Oid relid)
* the data file.
*/
static void
filePlanForeignScan(Oid foreigntableid,
PlannerInfo *root,
RelOptInfo *baserel)
fileGetForeignPaths(PlannerInfo *root,
RelOptInfo *baserel,
Oid foreigntableid)
{
char *filename;
List *options;
FileFdwPlanState *fdw_private = (FileFdwPlanState *) baserel->fdw_private;
Cost startup_cost;
Cost total_cost;
/* Fetch options --- we only need filename at this point */
fileGetOptions(foreigntableid, &filename, &options);
/* Estimate costs and update baserel->rows */
estimate_costs(root, baserel, filename,
/* Estimate costs */
estimate_costs(root, baserel, fdw_private,
&startup_cost, &total_cost);
/* Create a ForeignPath node and add it as only possible path */
@ -422,6 +468,37 @@ filePlanForeignScan(Oid foreigntableid,
*/
}
/*
* fileGetForeignPlan
* Create a ForeignScan plan node for scanning the foreign table
*/
static ForeignScan *
fileGetForeignPlan(PlannerInfo *root,
RelOptInfo *baserel,
Oid foreigntableid,
ForeignPath *best_path,
List *tlist,
List *scan_clauses)
{
Index scan_relid = baserel->relid;
/*
* We have no native ability to evaluate restriction clauses, so we just
* put all the scan_clauses into the plan node's qual list for the
* executor to check. So all we have to do here is strip RestrictInfo
* nodes from the clauses and ignore pseudoconstants (which will be
* handled elsewhere).
*/
scan_clauses = extract_actual_clauses(scan_clauses, false);
/* Create the ForeignScan node */
return make_foreignscan(tlist,
scan_clauses,
scan_relid,
NIL, /* no expressions to evaluate */
NIL); /* no private state either */
}
/*
* fileExplainForeignScan
* Produce extra output for EXPLAIN
@ -568,38 +645,38 @@ fileReScanForeignScan(ForeignScanState *node)
}
/*
* Estimate costs of scanning a foreign table.
* Estimate size of a foreign table.
*
* In addition to setting *startup_cost and *total_cost, this should
* update baserel->rows.
* The main result is returned in baserel->rows. We also set
* fdw_private->pages and fdw_private->ntuples for later use in the cost
* calculation.
*/
static void
estimate_costs(PlannerInfo *root, RelOptInfo *baserel,
const char *filename,
Cost *startup_cost, Cost *total_cost)
estimate_size(PlannerInfo *root, RelOptInfo *baserel,
FileFdwPlanState *fdw_private)
{
struct stat stat_buf;
BlockNumber pages;
int tuple_width;
double ntuples;
double nrows;
Cost run_cost = 0;
Cost cpu_per_tuple;
/*
* Get size of the file. It might not be there at plan time, though, in
* which case we have to use a default estimate.
*/
if (stat(filename, &stat_buf) < 0)
if (stat(fdw_private->filename, &stat_buf) < 0)
stat_buf.st_size = 10 * BLCKSZ;
/*
* Convert size to pages for use in I/O cost estimate below.
* Convert size to pages for use in I/O cost estimate later.
*/
pages = (stat_buf.st_size + (BLCKSZ - 1)) / BLCKSZ;
if (pages < 1)
pages = 1;
fdw_private->pages = pages;
/*
* Estimate the number of tuples in the file. We back into this estimate
* using the planner's idea of the relation width; which is bogus if not
@ -611,6 +688,8 @@ estimate_costs(PlannerInfo *root, RelOptInfo *baserel,
ntuples = clamp_row_est((double) stat_buf.st_size / (double) tuple_width);
fdw_private->ntuples = ntuples;
/*
* Now estimate the number of rows returned by the scan after applying the
* baserestrictinfo quals. This is pretty bogus too, since the planner
@ -627,12 +706,28 @@ estimate_costs(PlannerInfo *root, RelOptInfo *baserel,
/* Save the output-rows estimate for the planner */
baserel->rows = nrows;
}
/*
* Estimate costs of scanning a foreign table.
*
* Results are returned in *startup_cost and *total_cost.
*/
static void
estimate_costs(PlannerInfo *root, RelOptInfo *baserel,
FileFdwPlanState *fdw_private,
Cost *startup_cost, Cost *total_cost)
{
BlockNumber pages = fdw_private->pages;
double ntuples = fdw_private->ntuples;
Cost run_cost = 0;
Cost cpu_per_tuple;
/*
* Now estimate costs. We estimate costs almost the same way as
* cost_seqscan(), thus assuming that I/O costs are equivalent to a
* regular table file of the same size. However, we take per-tuple CPU
* costs as 10x of a seqscan, to account for the cost of parsing records.
* We estimate costs almost the same way as cost_seqscan(), thus assuming
* that I/O costs are equivalent to a regular table file of the same size.
* However, we take per-tuple CPU costs as 10x of a seqscan, to account
* for the cost of parsing records.
*/
run_cost += seq_page_cost * pages;