Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * tsm_system_time.c
4 : * support routines for SYSTEM_TIME tablesample method
5 : *
6 : * The desire here is to produce a random sample with as many rows as possible
7 : * in no more than the specified amount of time. 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 time dependence, this method is necessarily unrepeatable.
13 : * However, we do what we can to reduce surprising behavior by selecting
14 : * the sampling pattern just once per query, much as in tsm_system_rows.
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 : * contrib/tsm_system_time/tsm_system_time.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 "miscadmin.h"
32 : #include "optimizer/optimizer.h"
33 : #include "portability/instr_time.h"
34 : #include "utils/sampling.h"
35 : #include "utils/spccache.h"
36 :
37 1 : PG_MODULE_MAGIC_EXT(
38 : .name = "tsm_system_time",
39 : .version = PG_VERSION
40 : );
41 :
42 2 : PG_FUNCTION_INFO_V1(tsm_system_time_handler);
43 :
44 :
45 : /* Private state */
46 : typedef struct
47 : {
48 : uint32 seed; /* random seed */
49 : double millis; /* time limit for sampling */
50 : instr_time start_time; /* scan start time */
51 : OffsetNumber lt; /* last tuple returned from current block */
52 : BlockNumber doneblocks; /* number of already-scanned blocks */
53 : BlockNumber lb; /* last block visited */
54 : /* these three values are not changed during a rescan: */
55 : BlockNumber nblocks; /* number of blocks in relation */
56 : BlockNumber firstblock; /* first block to sample from */
57 : BlockNumber step; /* step size, or 0 if not set yet */
58 : } SystemTimeSamplerData;
59 :
60 : static void system_time_samplescangetsamplesize(PlannerInfo *root,
61 : RelOptInfo *baserel,
62 : List *paramexprs,
63 : BlockNumber *pages,
64 : double *tuples);
65 : static void system_time_initsamplescan(SampleScanState *node,
66 : int eflags);
67 : static void system_time_beginsamplescan(SampleScanState *node,
68 : Datum *params,
69 : int nparams,
70 : uint32 seed);
71 : static BlockNumber system_time_nextsampleblock(SampleScanState *node, BlockNumber nblocks);
72 : static OffsetNumber system_time_nextsampletuple(SampleScanState *node,
73 : BlockNumber blockno,
74 : OffsetNumber maxoffset);
75 : static uint32 random_relative_prime(uint32 n, pg_prng_state *randstate);
76 :
77 :
78 : /*
79 : * Create a TsmRoutine descriptor for the SYSTEM_TIME method.
80 : */
81 : Datum
82 41 : tsm_system_time_handler(PG_FUNCTION_ARGS)
83 : {
84 41 : TsmRoutine *tsm = makeNode(TsmRoutine);
85 :
86 41 : tsm->parameterTypes = list_make1_oid(FLOAT8OID);
87 :
88 : /* See notes at head of file */
89 41 : tsm->repeatable_across_queries = false;
90 41 : tsm->repeatable_across_scans = false;
91 :
92 41 : tsm->SampleScanGetSampleSize = system_time_samplescangetsamplesize;
93 41 : tsm->InitSampleScan = system_time_initsamplescan;
94 41 : tsm->BeginSampleScan = system_time_beginsamplescan;
95 41 : tsm->NextSampleBlock = system_time_nextsampleblock;
96 41 : tsm->NextSampleTuple = system_time_nextsampletuple;
97 41 : tsm->EndSampleScan = NULL;
98 :
99 41 : PG_RETURN_POINTER(tsm);
100 : }
101 :
102 : /*
103 : * Sample size estimation.
104 : */
105 : static void
106 9 : system_time_samplescangetsamplesize(PlannerInfo *root,
107 : RelOptInfo *baserel,
108 : List *paramexprs,
109 : BlockNumber *pages,
110 : double *tuples)
111 : {
112 : Node *limitnode;
113 : double millis;
114 : double spc_random_page_cost;
115 : double npages;
116 : double ntuples;
117 :
118 : /* Try to extract an estimate for the limit time spec */
119 9 : limitnode = (Node *) linitial(paramexprs);
120 9 : limitnode = estimate_expression_value(root, limitnode);
121 :
122 9 : if (IsA(limitnode, Const) &&
123 7 : !((Const *) limitnode)->constisnull)
124 : {
125 7 : millis = DatumGetFloat8(((Const *) limitnode)->constvalue);
126 7 : if (millis < 0 || isnan(millis))
127 : {
128 : /* Default millis if the value is bogus */
129 2 : millis = 1000;
130 : }
131 : }
132 : else
133 : {
134 : /* Default millis if we didn't obtain a non-null Const */
135 2 : millis = 1000;
136 : }
137 :
138 : /* Get the planner's idea of cost per page read */
139 9 : get_tablespace_page_costs(baserel->reltablespace,
140 : &spc_random_page_cost,
141 : NULL);
142 :
143 : /*
144 : * Estimate the number of pages we can read by assuming that the cost
145 : * figure is expressed in milliseconds. This is completely, unmistakably
146 : * bogus, but we have to do something to produce an estimate and there's
147 : * no better answer.
148 : */
149 9 : if (spc_random_page_cost > 0)
150 9 : npages = millis / spc_random_page_cost;
151 : else
152 0 : npages = millis; /* even more bogus, but whatcha gonna do? */
153 :
154 : /* Clamp to sane value */
155 9 : npages = clamp_row_est(Min((double) baserel->pages, npages));
156 :
157 9 : if (baserel->tuples > 0 && baserel->pages > 0)
158 9 : {
159 : /* Estimate number of tuples returned based on tuple density */
160 9 : double density = baserel->tuples / (double) baserel->pages;
161 :
162 9 : ntuples = npages * density;
163 : }
164 : else
165 : {
166 : /* For lack of data, assume one tuple per page */
167 0 : ntuples = npages;
168 : }
169 :
170 : /* Clamp to the estimated relation size */
171 9 : ntuples = clamp_row_est(Min(baserel->tuples, ntuples));
172 :
173 9 : *pages = npages;
174 9 : *tuples = ntuples;
175 9 : }
176 :
177 : /*
178 : * Initialize during executor setup.
179 : */
180 : static void
181 9 : system_time_initsamplescan(SampleScanState *node, int eflags)
182 : {
183 9 : node->tsm_state = palloc0_object(SystemTimeSamplerData);
184 : /* Note the above leaves tsm_state->step equal to zero */
185 9 : }
186 :
187 : /*
188 : * Examine parameters and prepare for a sample scan.
189 : */
190 : static void
191 6 : system_time_beginsamplescan(SampleScanState *node,
192 : Datum *params,
193 : int nparams,
194 : uint32 seed)
195 : {
196 6 : SystemTimeSamplerData *sampler = (SystemTimeSamplerData *) node->tsm_state;
197 6 : double millis = DatumGetFloat8(params[0]);
198 :
199 6 : if (millis < 0 || isnan(millis))
200 1 : ereport(ERROR,
201 : (errcode(ERRCODE_INVALID_TABLESAMPLE_ARGUMENT),
202 : errmsg("sample collection time must not be negative")));
203 :
204 5 : sampler->seed = seed;
205 5 : sampler->millis = millis;
206 5 : sampler->lt = InvalidOffsetNumber;
207 5 : sampler->doneblocks = 0;
208 : /* start_time, lb will be initialized during first NextSampleBlock call */
209 : /* we intentionally do not change nblocks/firstblock/step here */
210 5 : }
211 :
212 : /*
213 : * Select next block to sample.
214 : *
215 : * Uses linear probing algorithm for picking next block.
216 : */
217 : static BlockNumber
218 26 : system_time_nextsampleblock(SampleScanState *node, BlockNumber nblocks)
219 : {
220 26 : SystemTimeSamplerData *sampler = (SystemTimeSamplerData *) node->tsm_state;
221 : instr_time cur_time;
222 :
223 : /* First call within scan? */
224 26 : if (sampler->doneblocks == 0)
225 : {
226 : /* First scan within query? */
227 5 : if (sampler->step == 0)
228 : {
229 : /* Initialize now that we have scan descriptor */
230 : pg_prng_state randstate;
231 :
232 : /* If relation is empty, there's nothing to scan */
233 4 : if (nblocks == 0)
234 0 : return InvalidBlockNumber;
235 :
236 : /* We only need an RNG during this setup step */
237 4 : sampler_random_init_state(sampler->seed, &randstate);
238 :
239 : /* Compute nblocks/firstblock/step only once per query */
240 4 : sampler->nblocks = nblocks;
241 :
242 : /* Choose random starting block within the relation */
243 : /* (Actually this is the predecessor of the first block visited) */
244 4 : sampler->firstblock = sampler_random_fract(&randstate) *
245 4 : sampler->nblocks;
246 :
247 : /* Find relative prime as step size for linear probing */
248 4 : sampler->step = random_relative_prime(sampler->nblocks, &randstate);
249 : }
250 :
251 : /* Reinitialize lb and start_time */
252 5 : sampler->lb = sampler->firstblock;
253 5 : INSTR_TIME_SET_CURRENT(sampler->start_time);
254 : }
255 :
256 : /* If we've read all blocks in relation, we're done */
257 26 : if (++sampler->doneblocks > sampler->nblocks)
258 3 : return InvalidBlockNumber;
259 :
260 : /* If we've used up all the allotted time, we're done */
261 23 : INSTR_TIME_SET_CURRENT(cur_time);
262 23 : INSTR_TIME_SUBTRACT(cur_time, sampler->start_time);
263 23 : if (INSTR_TIME_GET_MILLISEC(cur_time) >= sampler->millis)
264 2 : return InvalidBlockNumber;
265 :
266 : /*
267 : * It's probably impossible for scan->rs_nblocks to decrease between scans
268 : * within a query; but just in case, loop until we select a block number
269 : * less than scan->rs_nblocks. We don't care if scan->rs_nblocks has
270 : * increased since the first scan.
271 : */
272 : do
273 : {
274 : /* Advance lb, using uint64 arithmetic to forestall overflow */
275 21 : sampler->lb = ((uint64) sampler->lb + sampler->step) % sampler->nblocks;
276 21 : } while (sampler->lb >= nblocks);
277 :
278 21 : return sampler->lb;
279 : }
280 :
281 : /*
282 : * Select next sampled tuple in current block.
283 : *
284 : * In block sampling, we just want to sample all the tuples in each selected
285 : * block.
286 : *
287 : * When we reach end of the block, return InvalidOffsetNumber which tells
288 : * SampleScan to go to next block.
289 : */
290 : static OffsetNumber
291 114 : system_time_nextsampletuple(SampleScanState *node,
292 : BlockNumber blockno,
293 : OffsetNumber maxoffset)
294 : {
295 114 : SystemTimeSamplerData *sampler = (SystemTimeSamplerData *) node->tsm_state;
296 114 : OffsetNumber tupoffset = sampler->lt;
297 :
298 : /* Advance to next possible offset on page */
299 114 : if (tupoffset == InvalidOffsetNumber)
300 21 : tupoffset = FirstOffsetNumber;
301 : else
302 93 : tupoffset++;
303 :
304 : /* Done? */
305 114 : if (tupoffset > maxoffset)
306 21 : tupoffset = InvalidOffsetNumber;
307 :
308 114 : sampler->lt = tupoffset;
309 :
310 114 : return tupoffset;
311 : }
312 :
313 : /*
314 : * Compute greatest common divisor of two uint32's.
315 : */
316 : static uint32
317 4 : gcd(uint32 a, uint32 b)
318 : {
319 : uint32 c;
320 :
321 13 : while (a != 0)
322 : {
323 9 : c = a;
324 9 : a = b % a;
325 9 : b = c;
326 : }
327 :
328 4 : return b;
329 : }
330 :
331 : /*
332 : * Pick a random value less than and relatively prime to n, if possible
333 : * (else return 1).
334 : */
335 : static uint32
336 4 : random_relative_prime(uint32 n, pg_prng_state *randstate)
337 : {
338 : uint32 r;
339 :
340 : /* Safety check to avoid infinite loop or zero result for small n. */
341 4 : if (n <= 1)
342 0 : return 1;
343 :
344 : /*
345 : * This should only take 2 or 3 iterations as the probability of 2 numbers
346 : * being relatively prime is ~61%; but just in case, we'll include a
347 : * CHECK_FOR_INTERRUPTS in the loop.
348 : */
349 : do
350 : {
351 4 : CHECK_FOR_INTERRUPTS();
352 4 : r = (uint32) (sampler_random_fract(randstate) * n);
353 4 : } while (r == 0 || gcd(r, n) > 1);
354 :
355 4 : return r;
356 : }
|
