Line data Source code
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) 1996-2026, 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
67 320 : tsm_system_handler(PG_FUNCTION_ARGS)
68 : {
69 320 : TsmRoutine *tsm = makeNode(TsmRoutine);
70 :
71 320 : tsm->parameterTypes = list_make1_oid(FLOAT4OID);
72 320 : tsm->repeatable_across_queries = true;
73 320 : tsm->repeatable_across_scans = true;
74 320 : tsm->SampleScanGetSampleSize = system_samplescangetsamplesize;
75 320 : tsm->InitSampleScan = system_initsamplescan;
76 320 : tsm->BeginSampleScan = system_beginsamplescan;
77 320 : tsm->NextSampleBlock = system_nextsampleblock;
78 320 : tsm->NextSampleTuple = system_nextsampletuple;
79 320 : tsm->EndSampleScan = NULL;
80 :
81 320 : PG_RETURN_POINTER(tsm);
82 : }
83 :
84 : /*
85 : * Sample size estimation.
86 : */
87 : static void
88 75 : 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 75 : pctnode = (Node *) linitial(paramexprs);
99 75 : pctnode = estimate_expression_value(root, pctnode);
100 :
101 75 : if (IsA(pctnode, Const) &&
102 42 : !((Const *) pctnode)->constisnull)
103 : {
104 39 : samplefract = DatumGetFloat4(((Const *) pctnode)->constvalue);
105 39 : if (samplefract >= 0 && samplefract <= 100 && !isnan(samplefract))
106 33 : samplefract /= 100.0f;
107 : else
108 : {
109 : /* Default samplefract if the value is bogus */
110 6 : samplefract = 0.1f;
111 : }
112 : }
113 : else
114 : {
115 : /* Default samplefract if we didn't obtain a non-null Const */
116 36 : samplefract = 0.1f;
117 : }
118 :
119 : /* We'll visit a sample of the pages ... */
120 75 : *pages = clamp_row_est(baserel->pages * samplefract);
121 :
122 : /* ... and hopefully get a representative number of tuples from them */
123 75 : *tuples = clamp_row_est(baserel->tuples * samplefract);
124 75 : }
125 :
126 : /*
127 : * Initialize during executor setup.
128 : */
129 : static void
130 75 : system_initsamplescan(SampleScanState *node, int eflags)
131 : {
132 75 : node->tsm_state = palloc0_object(SystemSamplerData);
133 75 : }
134 :
135 : /*
136 : * Examine parameters and prepare for a sample scan.
137 : */
138 : static void
139 42 : system_beginsamplescan(SampleScanState *node,
140 : Datum *params,
141 : int nparams,
142 : uint32 seed)
143 : {
144 42 : SystemSamplerData *sampler = (SystemSamplerData *) node->tsm_state;
145 42 : double percent = DatumGetFloat4(params[0]);
146 : double dcutoff;
147 :
148 42 : if (percent < 0 || percent > 100 || isnan(percent))
149 6 : 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 36 : dcutoff = rint(((double) PG_UINT32_MAX + 1) * percent / 100);
159 36 : sampler->cutoff = (uint64) dcutoff;
160 36 : sampler->seed = seed;
161 36 : sampler->nextblock = 0;
162 36 : 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 36 : node->use_bulkread = (percent >= 1);
171 36 : node->use_pagemode = true;
172 36 : }
173 :
174 : /*
175 : * Select next block to sample.
176 : */
177 : static BlockNumber
178 2163 : system_nextsampleblock(SampleScanState *node, BlockNumber nblocks)
179 : {
180 2163 : SystemSamplerData *sampler = (SystemSamplerData *) node->tsm_state;
181 2163 : 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 2163 : hashinput[1] = sampler->seed;
194 :
195 : /*
196 : * Loop over block numbers until finding suitable block or reaching end of
197 : * relation.
198 : */
199 4266 : for (; nextblock < nblocks; nextblock++)
200 : {
201 : uint32 hash;
202 :
203 4233 : hashinput[0] = nextblock;
204 :
205 4233 : hash = DatumGetUInt32(hash_any((const unsigned char *) hashinput,
206 : (int) sizeof(hashinput)));
207 4233 : if (hash < sampler->cutoff)
208 2130 : break;
209 : }
210 :
211 2163 : if (nextblock < nblocks)
212 : {
213 : /* Found a suitable block; remember where we should start next time */
214 2130 : sampler->nextblock = nextblock + 1;
215 2130 : return nextblock;
216 : }
217 :
218 : /* Done, but let's reset nextblock to 0 for safety. */
219 33 : sampler->nextblock = 0;
220 33 : 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 62289 : system_nextsampletuple(SampleScanState *node,
237 : BlockNumber blockno,
238 : OffsetNumber maxoffset)
239 : {
240 62289 : SystemSamplerData *sampler = (SystemSamplerData *) node->tsm_state;
241 62289 : OffsetNumber tupoffset = sampler->lt;
242 :
243 : /* Advance to next possible offset on page */
244 62289 : if (tupoffset == InvalidOffsetNumber)
245 2130 : tupoffset = FirstOffsetNumber;
246 : else
247 60159 : tupoffset++;
248 :
249 : /* Done? */
250 62289 : if (tupoffset > maxoffset)
251 2127 : tupoffset = InvalidOffsetNumber;
252 :
253 62289 : sampler->lt = tupoffset;
254 :
255 62289 : return tupoffset;
256 : }
|
