| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* system.c |
| 4 |
* support routines for SYSTEM tablesample method |
| 5 |
* |
| 6 |
* To ensure repeatability of samples, it is necessary that selection of a |
| 7 |
* given tuple be history-independent; otherwise syncscanning would break |
| 8 |
* repeatability, to say nothing of logically-irrelevant maintenance such |
| 9 |
* as physical extension or shortening of the relation. |
| 10 |
* |
| 11 |
* To achieve that, we proceed by hashing each candidate block number together |
| 12 |
* with the active seed, and then selecting it if the hash is less than the |
| 13 |
* cutoff value computed from the selection probability by BeginSampleScan. |
| 14 |
* |
| 15 |
* |
| 16 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 17 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 18 |
* |
| 19 |
* IDENTIFICATION |
| 20 |
* src/backend/access/tablesample/system.c |
| 21 |
* |
| 22 |
*------------------------------------------------------------------------- |
| 23 |
*/ |
| 24 |
|
| 25 |
#include "postgres.h" |
| 26 |
|
| 27 |
#include <math.h> |
| 28 |
|
| 29 |
#include "access/tsmapi.h" |
| 30 |
#include "catalog/pg_type.h" |
| 31 |
#include "common/hashfn.h" |
| 32 |
#include "optimizer/optimizer.h" |
| 33 |
#include "utils/fmgrprotos.h" |
| 34 |
|
| 35 |
|
| 36 |
/* Private state */ |
| 37 |
typedef struct |
| 38 |
{ |
| 39 |
uint64 cutoff; /* select blocks with hash less than this */ |
| 40 |
uint32 seed; /* random seed */ |
| 41 |
BlockNumber nextblock; /* next block to consider sampling */ |
| 42 |
OffsetNumber lt; /* last tuple returned from current block */ |
| 43 |
} SystemSamplerData; |
| 44 |
|
| 45 |
|
| 46 |
static void system_samplescangetsamplesize(PlannerInfo *root, |
| 47 |
RelOptInfo *baserel, |
| 48 |
List *paramexprs, |
| 49 |
BlockNumber *pages, |
| 50 |
double *tuples); |
| 51 |
static void system_initsamplescan(SampleScanState *node, |
| 52 |
int eflags); |
| 53 |
static void system_beginsamplescan(SampleScanState *node, |
| 54 |
Datum *params, |
| 55 |
int nparams, |
| 56 |
uint32 seed); |
| 57 |
static BlockNumber system_nextsampleblock(SampleScanState *node, BlockNumber nblocks); |
| 58 |
static OffsetNumber system_nextsampletuple(SampleScanState *node, |
| 59 |
BlockNumber blockno, |
| 60 |
OffsetNumber maxoffset); |
| 61 |
|
| 62 |
|
| 63 |
/* |
| 64 |
* Create a TsmRoutine descriptor for the SYSTEM method. |
| 65 |
*/ |
| 66 |
Datum |
| 67 |
tsm_system_handler(PG_FUNCTION_ARGS) |
| 68 |
{ |
| 69 |
TsmRoutine *tsm = makeNode(TsmRoutine); |
| 70 |
|
| 71 |
tsm->parameterTypes = list_make1_oid(FLOAT4OID); |
| 72 |
tsm->repeatable_across_queries = true; |
| 73 |
tsm->repeatable_across_scans = true; |
| 74 |
tsm->SampleScanGetSampleSize = system_samplescangetsamplesize; |
| 75 |
tsm->InitSampleScan = system_initsamplescan; |
| 76 |
tsm->BeginSampleScan = system_beginsamplescan; |
| 77 |
tsm->NextSampleBlock = system_nextsampleblock; |
| 78 |
tsm->NextSampleTuple = system_nextsampletuple; |
| 79 |
tsm->EndSampleScan = NULL; |
| 80 |
|
| 81 |
PG_RETURN_POINTER(tsm); |
| 82 |
} |
| 83 |
|
| 84 |
/* |
| 85 |
* Sample size estimation. |
| 86 |
*/ |
| 87 |
static void |
| 88 |
system_samplescangetsamplesize(PlannerInfo *root, |
| 89 |
RelOptInfo *baserel, |
| 90 |
List *paramexprs, |
| 91 |
BlockNumber *pages, |
| 92 |
double *tuples) |
| 93 |
{ |
| 94 |
Node *pctnode; |
| 95 |
float4 samplefract; |
| 96 |
|
| 97 |
/* Try to extract an estimate for the sample percentage */ |
| 98 |
pctnode = (Node *) linitial(paramexprs); |
| 99 |
pctnode = estimate_expression_value(root, pctnode); |
| 100 |
|
| 101 |
if (IsA(pctnode, Const) && |
| 102 |
!((Const *) pctnode)->constisnull) |
| 103 |
{ |
| 104 |
samplefract = DatumGetFloat4(((Const *) pctnode)->constvalue); |
| 105 |
if (samplefract >= 0 && samplefract <= 100 && !isnan(samplefract)) |
| 106 |
samplefract /= 100.0f; |
| 107 |
else |
| 108 |
{ |
| 109 |
/* Default samplefract if the value is bogus */ |
| 110 |
samplefract = 0.1f; |
| 111 |
} |
| 112 |
} |
| 113 |
else |
| 114 |
{ |
| 115 |
/* Default samplefract if we didn't obtain a non-null Const */ |
| 116 |
samplefract = 0.1f; |
| 117 |
} |
| 118 |
|
| 119 |
/* We'll visit a sample of the pages ... */ |
| 120 |
*pages = clamp_row_est(baserel->pages * samplefract); |
| 121 |
|
| 122 |
/* ... and hopefully get a representative number of tuples from them */ |
| 123 |
*tuples = clamp_row_est(baserel->tuples * samplefract); |
| 124 |
} |
| 125 |
|
| 126 |
/* |
| 127 |
* Initialize during executor setup. |
| 128 |
*/ |
| 129 |
static void |
| 130 |
system_initsamplescan(SampleScanState *node, int eflags) |
| 131 |
{ |
| 132 |
node->tsm_state = palloc0(sizeof(SystemSamplerData)); |
| 133 |
} |
| 134 |
|
| 135 |
/* |
| 136 |
* Examine parameters and prepare for a sample scan. |
| 137 |
*/ |
| 138 |
static void |
| 139 |
system_beginsamplescan(SampleScanState *node, |
| 140 |
Datum *params, |
| 141 |
int nparams, |
| 142 |
uint32 seed) |
| 143 |
{ |
| 144 |
SystemSamplerData *sampler = (SystemSamplerData *) node->tsm_state; |
| 145 |
double percent = DatumGetFloat4(params[0]); |
| 146 |
double dcutoff; |
| 147 |
|
| 148 |
if (percent < 0 || percent > 100 || isnan(percent)) |
| 149 |
ereport(ERROR, |
| 150 |
(errcode(ERRCODE_INVALID_TABLESAMPLE_ARGUMENT), |
| 151 |
errmsg("sample percentage must be between 0 and 100"))); |
| 152 |
|
| 153 |
/* |
| 154 |
* The cutoff is sample probability times (PG_UINT32_MAX + 1); we have to |
| 155 |
* store that as a uint64, of course. Note that this gives strictly |
| 156 |
* correct behavior at the limits of zero or one probability. |
| 157 |
*/ |
| 158 |
dcutoff = rint(((double) PG_UINT32_MAX + 1) * percent / 100); |
| 159 |
sampler->cutoff = (uint64) dcutoff; |
| 160 |
sampler->seed = seed; |
| 161 |
sampler->nextblock = 0; |
| 162 |
sampler->lt = InvalidOffsetNumber; |
| 163 |
|
| 164 |
/* |
| 165 |
* Bulkread buffer access strategy probably makes sense unless we're |
| 166 |
* scanning a very small fraction of the table. The 1% cutoff here is a |
| 167 |
* guess. We should use pagemode visibility checking, since we scan all |
| 168 |
* tuples on each selected page. |
| 169 |
*/ |
| 170 |
node->use_bulkread = (percent >= 1); |
| 171 |
node->use_pagemode = true; |
| 172 |
} |
| 173 |
|
| 174 |
/* |
| 175 |
* Select next block to sample. |
| 176 |
*/ |
| 177 |
static BlockNumber |
| 178 |
system_nextsampleblock(SampleScanState *node, BlockNumber nblocks) |
| 179 |
{ |
| 180 |
SystemSamplerData *sampler = (SystemSamplerData *) node->tsm_state; |
| 181 |
BlockNumber nextblock = sampler->nextblock; |
| 182 |
uint32 hashinput[2]; |
| 183 |
|
| 184 |
/* |
| 185 |
* We compute the hash by applying hash_any to an array of 2 uint32's |
| 186 |
* containing the block number and seed. This is efficient to set up, and |
| 187 |
* with the current implementation of hash_any, it gives |
| 188 |
* machine-independent results, which is a nice property for regression |
| 189 |
* testing. |
| 190 |
* |
| 191 |
* These words in the hash input are the same throughout the block: |
| 192 |
*/ |
| 193 |
hashinput[1] = sampler->seed; |
| 194 |
|
| 195 |
/* |
| 196 |
* Loop over block numbers until finding suitable block or reaching end of |
| 197 |
* relation. |
| 198 |
*/ |
| 199 |
for (; nextblock < nblocks; nextblock++) |
| 200 |
{ |
| 201 |
uint32 hash; |
| 202 |
|
| 203 |
hashinput[0] = nextblock; |
| 204 |
|
| 205 |
hash = DatumGetUInt32(hash_any((const unsigned char *) hashinput, |
| 206 |
(int) sizeof(hashinput))); |
| 207 |
if (hash < sampler->cutoff) |
| 208 |
break; |
| 209 |
} |
| 210 |
|
| 211 |
if (nextblock < nblocks) |
| 212 |
{ |
| 213 |
/* Found a suitable block; remember where we should start next time */ |
| 214 |
sampler->nextblock = nextblock + 1; |
| 215 |
return nextblock; |
| 216 |
} |
| 217 |
|
| 218 |
/* Done, but let's reset nextblock to 0 for safety. */ |
| 219 |
sampler->nextblock = 0; |
| 220 |
return InvalidBlockNumber; |
| 221 |
} |
| 222 |
|
| 223 |
/* |
| 224 |
* Select next sampled tuple in current block. |
| 225 |
* |
| 226 |
* In block sampling, we just want to sample all the tuples in each selected |
| 227 |
* block. |
| 228 |
* |
| 229 |
* It is OK here to return an offset without knowing if the tuple is visible |
| 230 |
* (or even exists); nodeSamplescan.c will deal with that. |
| 231 |
* |
| 232 |
* When we reach end of the block, return InvalidOffsetNumber which tells |
| 233 |
* SampleScan to go to next block. |
| 234 |
*/ |
| 235 |
static OffsetNumber |
| 236 |
system_nextsampletuple(SampleScanState *node, |
| 237 |
BlockNumber blockno, |
| 238 |
OffsetNumber maxoffset) |
| 239 |
{ |
| 240 |
SystemSamplerData *sampler = (SystemSamplerData *) node->tsm_state; |
| 241 |
OffsetNumber tupoffset = sampler->lt; |
| 242 |
|
| 243 |
/* Advance to next possible offset on page */ |
| 244 |
if (tupoffset == InvalidOffsetNumber) |
| 245 |
tupoffset = FirstOffsetNumber; |
| 246 |
else |
| 247 |
tupoffset++; |
| 248 |
|
| 249 |
/* Done? */ |
| 250 |
if (tupoffset > maxoffset) |
| 251 |
tupoffset = InvalidOffsetNumber; |
| 252 |
|
| 253 |
sampler->lt = tupoffset; |
| 254 |
|
| 255 |
return tupoffset; |
| 256 |
} |