Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * tsm_system_rows.c
4 : : * support routines for SYSTEM_ROWS tablesample method
5 : : *
6 : : * The desire here is to produce a random sample with a given number of rows
7 : : * (or the whole relation, if that is fewer rows). We use a block-sampling
8 : : * approach. To ensure that the whole relation will be visited if necessary,
9 : : * we start at a randomly chosen block and then advance with a stride that
10 : : * is randomly chosen but is relatively prime to the relation's nblocks.
11 : : *
12 : : * Because of the dependence on nblocks, this method cannot be repeatable
13 : : * across queries. (Even if the user hasn't explicitly changed the relation,
14 : : * maintenance activities such as autovacuum might change nblocks.) However,
15 : : * we can at least make it repeatable across scans, by determining the
16 : : * sampling pattern only once on the first scan. This means that rescans
17 : : * won't visit blocks added after the first scan, but that is fine since
18 : : * such blocks shouldn't contain any visible tuples anyway.
19 : : *
20 : : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
21 : : * Portions Copyright (c) 1994, Regents of the University of California
22 : : *
23 : : * IDENTIFICATION
24 : : * contrib/tsm_system_rows/tsm_system_rows.c
25 : : *
26 : : *-------------------------------------------------------------------------
27 : : */
28 : :
29 : : #include "postgres.h"
30 : :
31 : : #include "access/tsmapi.h"
32 : : #include "catalog/pg_type.h"
33 : : #include "miscadmin.h"
34 : : #include "optimizer/optimizer.h"
35 : : #include "utils/sampling.h"
36 : :
164 tgl@sss.pgh.pa.us 37 :CBC 1 : PG_MODULE_MAGIC_EXT(
38 : : .name = "tsm_system_rows",
39 : : .version = PG_VERSION
40 : : );
41 : :
3696 42 : 2 : PG_FUNCTION_INFO_V1(tsm_system_rows_handler);
43 : :
44 : :
45 : : /* Private state */
46 : : typedef struct
47 : : {
48 : : uint32 seed; /* random seed */
49 : : int64 ntuples; /* number of tuples to return */
50 : : OffsetNumber lt; /* last tuple returned from current block */
51 : : BlockNumber doneblocks; /* number of already-scanned blocks */
52 : : BlockNumber lb; /* last block visited */
53 : : /* these three values are not changed during a rescan: */
54 : : BlockNumber nblocks; /* number of blocks in relation */
55 : : BlockNumber firstblock; /* first block to sample from */
56 : : BlockNumber step; /* step size, or 0 if not set yet */
57 : : } SystemRowsSamplerData;
58 : :
59 : : static void system_rows_samplescangetsamplesize(PlannerInfo *root,
60 : : RelOptInfo *baserel,
61 : : List *paramexprs,
62 : : BlockNumber *pages,
63 : : double *tuples);
64 : : static void system_rows_initsamplescan(SampleScanState *node,
65 : : int eflags);
66 : : static void system_rows_beginsamplescan(SampleScanState *node,
67 : : Datum *params,
68 : : int nparams,
69 : : uint32 seed);
70 : : static BlockNumber system_rows_nextsampleblock(SampleScanState *node, BlockNumber nblocks);
71 : : static OffsetNumber system_rows_nextsampletuple(SampleScanState *node,
72 : : BlockNumber blockno,
73 : : OffsetNumber maxoffset);
74 : : static uint32 random_relative_prime(uint32 n, pg_prng_state *randstate);
75 : :
76 : :
77 : : /*
78 : : * Create a TsmRoutine descriptor for the SYSTEM_ROWS method.
79 : : */
80 : : Datum
81 : 40 : tsm_system_rows_handler(PG_FUNCTION_ARGS)
82 : : {
83 : 40 : TsmRoutine *tsm = makeNode(TsmRoutine);
84 : :
85 : 40 : tsm->parameterTypes = list_make1_oid(INT8OID);
86 : :
87 : : /* See notes at head of file */
88 : 40 : tsm->repeatable_across_queries = false;
89 : 40 : tsm->repeatable_across_scans = true;
90 : :
91 : 40 : tsm->SampleScanGetSampleSize = system_rows_samplescangetsamplesize;
92 : 40 : tsm->InitSampleScan = system_rows_initsamplescan;
93 : 40 : tsm->BeginSampleScan = system_rows_beginsamplescan;
94 : 40 : tsm->NextSampleBlock = system_rows_nextsampleblock;
95 : 40 : tsm->NextSampleTuple = system_rows_nextsampletuple;
96 : 40 : tsm->EndSampleScan = NULL;
97 : :
98 : 40 : PG_RETURN_POINTER(tsm);
99 : : }
100 : :
101 : : /*
102 : : * Sample size estimation.
103 : : */
104 : : static void
105 : 9 : system_rows_samplescangetsamplesize(PlannerInfo *root,
106 : : RelOptInfo *baserel,
107 : : List *paramexprs,
108 : : BlockNumber *pages,
109 : : double *tuples)
110 : : {
111 : : Node *limitnode;
112 : : int64 ntuples;
113 : : double npages;
114 : :
115 : : /* Try to extract an estimate for the limit rowcount */
116 : 9 : limitnode = (Node *) linitial(paramexprs);
117 : 9 : limitnode = estimate_expression_value(root, limitnode);
118 : :
119 [ + + ]: 9 : if (IsA(limitnode, Const) &&
120 [ + - ]: 7 : !((Const *) limitnode)->constisnull)
121 : : {
122 : 7 : ntuples = DatumGetInt64(((Const *) limitnode)->constvalue);
123 [ + + ]: 7 : if (ntuples < 0)
124 : : {
125 : : /* Default ntuples if the value is bogus */
126 : 2 : ntuples = 1000;
127 : : }
128 : : }
129 : : else
130 : : {
131 : : /* Default ntuples if we didn't obtain a non-null Const */
132 : 2 : ntuples = 1000;
133 : : }
134 : :
135 : : /* Clamp to the estimated relation size */
136 [ + + ]: 9 : if (ntuples > baserel->tuples)
137 : 5 : ntuples = (int64) baserel->tuples;
138 : 9 : ntuples = clamp_row_est(ntuples);
139 : :
140 [ + - + - ]: 9 : if (baserel->tuples > 0 && baserel->pages > 0)
141 : 9 : {
142 : : /* Estimate number of pages visited based on tuple density */
143 : 9 : double density = baserel->tuples / (double) baserel->pages;
144 : :
145 : 9 : npages = ntuples / density;
146 : : }
147 : : else
148 : : {
149 : : /* For lack of data, assume one tuple per page */
3696 tgl@sss.pgh.pa.us 150 :UBC 0 : npages = ntuples;
151 : : }
152 : :
153 : : /* Clamp to sane value */
3696 tgl@sss.pgh.pa.us 154 [ - + ]:CBC 9 : npages = clamp_row_est(Min((double) baserel->pages, npages));
155 : :
156 : 9 : *pages = npages;
157 : 9 : *tuples = ntuples;
3767 simon@2ndQuadrant.co 158 : 9 : }
159 : :
160 : : /*
161 : : * Initialize during executor setup.
162 : : */
163 : : static void
3696 tgl@sss.pgh.pa.us 164 : 9 : system_rows_initsamplescan(SampleScanState *node, int eflags)
165 : : {
166 : 9 : node->tsm_state = palloc0(sizeof(SystemRowsSamplerData));
167 : : /* Note the above leaves tsm_state->step equal to zero */
3767 simon@2ndQuadrant.co 168 : 9 : }
169 : :
170 : : /*
171 : : * Examine parameters and prepare for a sample scan.
172 : : */
173 : : static void
3696 tgl@sss.pgh.pa.us 174 : 9 : system_rows_beginsamplescan(SampleScanState *node,
175 : : Datum *params,
176 : : int nparams,
177 : : uint32 seed)
178 : : {
179 : 9 : SystemRowsSamplerData *sampler = (SystemRowsSamplerData *) node->tsm_state;
180 : 9 : int64 ntuples = DatumGetInt64(params[0]);
181 : :
182 [ + + ]: 9 : if (ntuples < 0)
183 [ + - ]: 1 : ereport(ERROR,
184 : : (errcode(ERRCODE_INVALID_TABLESAMPLE_ARGUMENT),
185 : : errmsg("sample size must not be negative")));
186 : :
187 : 8 : sampler->seed = seed;
188 : 8 : sampler->ntuples = ntuples;
189 : 8 : sampler->lt = InvalidOffsetNumber;
190 : 8 : sampler->doneblocks = 0;
191 : : /* lb will be initialized during first NextSampleBlock call */
192 : : /* we intentionally do not change nblocks/firstblock/step here */
193 : :
194 : : /*
195 : : * We *must* use pagemode visibility checking in this module, so force
196 : : * that even though it's currently default.
197 : : */
198 : 8 : node->use_pagemode = true;
3767 simon@2ndQuadrant.co 199 : 8 : }
200 : :
201 : : /*
202 : : * Select next block to sample.
203 : : *
204 : : * Uses linear probing algorithm for picking next block.
205 : : */
206 : : static BlockNumber
2352 andres@anarazel.de 207 : 32 : system_rows_nextsampleblock(SampleScanState *node, BlockNumber nblocks)
208 : : {
3696 tgl@sss.pgh.pa.us 209 : 32 : SystemRowsSamplerData *sampler = (SystemRowsSamplerData *) node->tsm_state;
210 : :
211 : : /* First call within scan? */
212 [ + + ]: 32 : if (sampler->doneblocks == 0)
213 : : {
214 : : /* First scan within query? */
215 [ + + ]: 8 : if (sampler->step == 0)
216 : : {
217 : : /* Initialize now that we have scan descriptor */
218 : : pg_prng_state randstate;
219 : :
220 : : /* If relation is empty, there's nothing to scan */
2352 andres@anarazel.de 221 [ - + ]: 6 : if (nblocks == 0)
3696 tgl@sss.pgh.pa.us 222 :UBC 0 : return InvalidBlockNumber;
223 : :
224 : : /* We only need an RNG during this setup step */
1378 tgl@sss.pgh.pa.us 225 :CBC 6 : sampler_random_init_state(sampler->seed, &randstate);
226 : :
227 : : /* Compute nblocks/firstblock/step only once per query */
2352 andres@anarazel.de 228 : 6 : sampler->nblocks = nblocks;
229 : :
230 : : /* Choose random starting block within the relation */
231 : : /* (Actually this is the predecessor of the first block visited) */
1378 tgl@sss.pgh.pa.us 232 : 6 : sampler->firstblock = sampler_random_fract(&randstate) *
3696 233 : 6 : sampler->nblocks;
234 : :
235 : : /* Find relative prime as step size for linear probing */
1378 236 : 6 : sampler->step = random_relative_prime(sampler->nblocks, &randstate);
237 : : }
238 : :
239 : : /* Reinitialize lb */
3696 240 : 8 : sampler->lb = sampler->firstblock;
241 : : }
242 : :
243 : : /* If we've read all blocks or returned all needed tuples, we're done */
244 [ + + ]: 32 : if (++sampler->doneblocks > sampler->nblocks ||
2352 andres@anarazel.de 245 [ + + ]: 30 : node->donetuples >= sampler->ntuples)
3696 tgl@sss.pgh.pa.us 246 : 8 : return InvalidBlockNumber;
247 : :
248 : : /*
249 : : * It's probably impossible for scan->rs_nblocks to decrease between scans
250 : : * within a query; but just in case, loop until we select a block number
251 : : * less than scan->rs_nblocks. We don't care if scan->rs_nblocks has
252 : : * increased since the first scan.
253 : : */
254 : : do
255 : : {
256 : : /* Advance lb, using uint64 arithmetic to forestall overflow */
257 : 24 : sampler->lb = ((uint64) sampler->lb + sampler->step) % sampler->nblocks;
2352 andres@anarazel.de 258 [ - + ]: 24 : } while (sampler->lb >= nblocks);
259 : :
3696 tgl@sss.pgh.pa.us 260 : 24 : return sampler->lb;
261 : : }
262 : :
263 : : /*
264 : : * Select next sampled tuple in current block.
265 : : *
266 : : * In block sampling, we just want to sample all the tuples in each selected
267 : : * block.
268 : : *
269 : : * When we reach end of the block, return InvalidOffsetNumber which tells
270 : : * SampleScan to go to next block.
271 : : */
272 : : static OffsetNumber
273 : 127 : system_rows_nextsampletuple(SampleScanState *node,
274 : : BlockNumber blockno,
275 : : OffsetNumber maxoffset)
276 : : {
277 : 127 : SystemRowsSamplerData *sampler = (SystemRowsSamplerData *) node->tsm_state;
278 : 127 : OffsetNumber tupoffset = sampler->lt;
279 : :
280 : : /* Quit if we've returned all needed tuples */
2352 andres@anarazel.de 281 [ + + ]: 127 : if (node->donetuples >= sampler->ntuples)
3696 tgl@sss.pgh.pa.us 282 : 4 : return InvalidOffsetNumber;
283 : :
284 : : /* Advance to next possible offset on page */
2352 andres@anarazel.de 285 [ + + ]: 123 : if (tupoffset == InvalidOffsetNumber)
286 : 24 : tupoffset = FirstOffsetNumber;
287 : : else
288 : 99 : tupoffset++;
289 : :
290 : : /* Done? */
291 [ + + ]: 123 : if (tupoffset > maxoffset)
292 : 20 : tupoffset = InvalidOffsetNumber;
293 : :
3696 tgl@sss.pgh.pa.us 294 : 123 : sampler->lt = tupoffset;
295 : :
296 : 123 : return tupoffset;
297 : : }
298 : :
299 : : /*
300 : : * Compute greatest common divisor of two uint32's.
301 : : */
302 : : static uint32
3759 bruce@momjian.us 303 : 6 : gcd(uint32 a, uint32 b)
304 : : {
305 : : uint32 c;
306 : :
3767 simon@2ndQuadrant.co 307 [ + + ]: 21 : while (a != 0)
308 : : {
309 : 15 : c = a;
310 : 15 : a = b % a;
311 : 15 : b = c;
312 : : }
313 : :
314 : 6 : return b;
315 : : }
316 : :
317 : : /*
318 : : * Pick a random value less than and relatively prime to n, if possible
319 : : * (else return 1).
320 : : */
321 : : static uint32
1378 tgl@sss.pgh.pa.us 322 : 6 : random_relative_prime(uint32 n, pg_prng_state *randstate)
323 : : {
324 : : uint32 r;
325 : :
326 : : /* Safety check to avoid infinite loop or zero result for small n. */
3696 327 [ - + ]: 6 : if (n <= 1)
3696 tgl@sss.pgh.pa.us 328 :UBC 0 : return 1;
329 : :
330 : : /*
331 : : * This should only take 2 or 3 iterations as the probability of 2 numbers
332 : : * being relatively prime is ~61%; but just in case, we'll include a
333 : : * CHECK_FOR_INTERRUPTS in the loop.
334 : : */
335 : : do
336 : : {
3767 simon@2ndQuadrant.co 337 [ - + ]:CBC 6 : CHECK_FOR_INTERRUPTS();
3696 tgl@sss.pgh.pa.us 338 : 6 : r = (uint32) (sampler_random_fract(randstate) * n);
339 [ - + - + ]: 6 : } while (r == 0 || gcd(r, n) > 1);
340 : :
3767 simon@2ndQuadrant.co 341 : 6 : return r;
342 : : }
|
