Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * rangetypes_typanalyze.c
4 : * Functions for gathering statistics from range columns
5 : *
6 : * For a range type column, histograms of lower and upper bounds, and
7 : * the fraction of NULL and empty ranges are collected.
8 : *
9 : * Both histograms have the same length, and they are combined into a
10 : * single array of ranges. This has the same shape as the histogram that
11 : * std_typanalyze would collect, but the values are different. Each range
12 : * in the array is a valid range, even though the lower and upper bounds
13 : * come from different tuples. In theory, the standard scalar selectivity
14 : * functions could be used with the combined histogram.
15 : *
16 : * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
17 : * Portions Copyright (c) 1994, Regents of the University of California
18 : *
19 : *
20 : * IDENTIFICATION
21 : * src/backend/utils/adt/rangetypes_typanalyze.c
22 : *
23 : *-------------------------------------------------------------------------
24 : */
25 : #include "postgres.h"
26 :
27 : #include "catalog/pg_operator.h"
28 : #include "commands/vacuum.h"
29 : #include "utils/float.h"
30 : #include "utils/fmgrprotos.h"
31 : #include "utils/lsyscache.h"
32 : #include "utils/rangetypes.h"
33 : #include "utils/multirangetypes.h"
34 : #include "varatt.h"
35 :
36 : static int float8_qsort_cmp(const void *a1, const void *a2, void *arg);
37 : static int range_bound_qsort_cmp(const void *a1, const void *a2, void *arg);
38 : static void compute_range_stats(VacAttrStats *stats,
39 : AnalyzeAttrFetchFunc fetchfunc, int samplerows,
40 : double totalrows);
41 :
42 : /*
43 : * range_typanalyze -- typanalyze function for range columns
44 : */
45 : Datum
46 20 : range_typanalyze(PG_FUNCTION_ARGS)
47 : {
48 20 : VacAttrStats *stats = (VacAttrStats *) PG_GETARG_POINTER(0);
49 : TypeCacheEntry *typcache;
50 20 : Form_pg_attribute attr = stats->attr;
51 :
52 : /* Get information about range type; note column might be a domain */
53 20 : typcache = range_get_typcache(fcinfo, getBaseType(stats->attrtypid));
54 :
55 20 : if (attr->attstattarget < 0)
56 20 : attr->attstattarget = default_statistics_target;
57 :
58 20 : stats->compute_stats = compute_range_stats;
59 20 : stats->extra_data = typcache;
60 : /* same as in std_typanalyze */
61 20 : stats->minrows = 300 * attr->attstattarget;
62 :
63 20 : PG_RETURN_BOOL(true);
64 : }
65 :
66 : /*
67 : * multirange_typanalyze -- typanalyze function for multirange columns
68 : *
69 : * We do the same analysis as for ranges, but on the smallest range that
70 : * completely includes the multirange.
71 : */
72 : Datum
73 12 : multirange_typanalyze(PG_FUNCTION_ARGS)
74 : {
75 12 : VacAttrStats *stats = (VacAttrStats *) PG_GETARG_POINTER(0);
76 : TypeCacheEntry *typcache;
77 12 : Form_pg_attribute attr = stats->attr;
78 :
79 : /* Get information about multirange type; note column might be a domain */
80 12 : typcache = multirange_get_typcache(fcinfo, getBaseType(stats->attrtypid));
81 :
82 12 : if (attr->attstattarget < 0)
83 12 : attr->attstattarget = default_statistics_target;
84 :
85 12 : stats->compute_stats = compute_range_stats;
86 12 : stats->extra_data = typcache;
87 : /* same as in std_typanalyze */
88 12 : stats->minrows = 300 * attr->attstattarget;
89 :
90 12 : PG_RETURN_BOOL(true);
91 : }
92 :
93 : /*
94 : * Comparison function for sorting float8s, used for range lengths.
95 : */
96 : static int
97 153880 : float8_qsort_cmp(const void *a1, const void *a2, void *arg)
98 : {
99 153880 : const float8 *f1 = (const float8 *) a1;
100 153880 : const float8 *f2 = (const float8 *) a2;
101 :
102 153880 : if (*f1 < *f2)
103 92 : return -1;
104 153788 : else if (*f1 == *f2)
105 136886 : return 0;
106 : else
107 16902 : return 1;
108 : }
109 :
110 : /*
111 : * Comparison function for sorting RangeBounds.
112 : */
113 : static int
114 2452334 : range_bound_qsort_cmp(const void *a1, const void *a2, void *arg)
115 : {
116 2452334 : RangeBound *b1 = (RangeBound *) a1;
117 2452334 : RangeBound *b2 = (RangeBound *) a2;
118 2452334 : TypeCacheEntry *typcache = (TypeCacheEntry *) arg;
119 :
120 2452334 : return range_cmp_bounds(typcache, b1, b2);
121 : }
122 :
123 : /*
124 : * compute_range_stats() -- compute statistics for a range column
125 : */
126 : static void
127 32 : compute_range_stats(VacAttrStats *stats, AnalyzeAttrFetchFunc fetchfunc,
128 : int samplerows, double totalrows)
129 : {
130 32 : TypeCacheEntry *typcache = (TypeCacheEntry *) stats->extra_data;
131 32 : TypeCacheEntry *mltrng_typcache = NULL;
132 : bool has_subdiff;
133 32 : int null_cnt = 0;
134 32 : int non_null_cnt = 0;
135 32 : int non_empty_cnt = 0;
136 32 : int empty_cnt = 0;
137 : int range_no;
138 : int slot_idx;
139 32 : int num_bins = stats->attr->attstattarget;
140 : int num_hist;
141 : float8 *lengths;
142 : RangeBound *lowers,
143 : *uppers;
144 32 : double total_width = 0;
145 :
146 32 : if (typcache->typtype == TYPTYPE_MULTIRANGE)
147 : {
148 12 : mltrng_typcache = typcache;
149 12 : typcache = typcache->rngtype;
150 : }
151 : else
152 : Assert(typcache->typtype == TYPTYPE_RANGE);
153 32 : has_subdiff = OidIsValid(typcache->rng_subdiff_finfo.fn_oid);
154 :
155 : /* Allocate memory to hold range bounds and lengths of the sample ranges. */
156 32 : lowers = (RangeBound *) palloc(sizeof(RangeBound) * samplerows);
157 32 : uppers = (RangeBound *) palloc(sizeof(RangeBound) * samplerows);
158 32 : lengths = (float8 *) palloc(sizeof(float8) * samplerows);
159 :
160 : /* Loop over the sample ranges. */
161 109134 : for (range_no = 0; range_no < samplerows; range_no++)
162 : {
163 : Datum value;
164 : bool isnull,
165 : empty;
166 : MultirangeType *multirange;
167 : RangeType *range;
168 : RangeBound lower,
169 : upper;
170 : float8 length;
171 :
172 109102 : vacuum_delay_point();
173 :
174 109102 : value = fetchfunc(stats, range_no, &isnull);
175 109102 : if (isnull)
176 : {
177 : /* range is null, just count that */
178 0 : null_cnt++;
179 0 : continue;
180 : }
181 :
182 : /*
183 : * XXX: should we ignore wide values, like std_typanalyze does, to
184 : * avoid bloating the statistics table?
185 : */
186 109102 : total_width += VARSIZE_ANY(DatumGetPointer(value));
187 :
188 : /* Get range and deserialize it for further analysis. */
189 109102 : if (mltrng_typcache != NULL)
190 : {
191 : /* Treat multiranges like a big range without gaps. */
192 22266 : multirange = DatumGetMultirangeTypeP(value);
193 22266 : if (!MultirangeIsEmpty(multirange))
194 : {
195 : RangeBound tmp;
196 :
197 19242 : multirange_get_bounds(typcache, multirange, 0,
198 : &lower, &tmp);
199 19242 : multirange_get_bounds(typcache, multirange,
200 19242 : multirange->rangeCount - 1,
201 : &tmp, &upper);
202 19242 : empty = false;
203 : }
204 : else
205 : {
206 3024 : empty = true;
207 : }
208 : }
209 : else
210 : {
211 86836 : range = DatumGetRangeTypeP(value);
212 86836 : range_deserialize(typcache, range, &lower, &upper, &empty);
213 : }
214 :
215 109102 : if (!empty)
216 : {
217 : /* Remember bounds and length for further usage in histograms */
218 92072 : lowers[non_empty_cnt] = lower;
219 92072 : uppers[non_empty_cnt] = upper;
220 :
221 92072 : if (lower.infinite || upper.infinite)
222 : {
223 : /* Length of any kind of an infinite range is infinite */
224 4042 : length = get_float8_infinity();
225 : }
226 88030 : else if (has_subdiff)
227 : {
228 : /*
229 : * For an ordinary range, use subdiff function between upper
230 : * and lower bound values.
231 : */
232 88030 : length = DatumGetFloat8(FunctionCall2Coll(&typcache->rng_subdiff_finfo,
233 : typcache->rng_collation,
234 : upper.val, lower.val));
235 : }
236 : else
237 : {
238 : /* Use default value of 1.0 if no subdiff is available. */
239 0 : length = 1.0;
240 : }
241 92072 : lengths[non_empty_cnt] = length;
242 :
243 92072 : non_empty_cnt++;
244 : }
245 : else
246 17030 : empty_cnt++;
247 :
248 109102 : non_null_cnt++;
249 : }
250 :
251 32 : slot_idx = 0;
252 :
253 : /* We can only compute real stats if we found some non-null values. */
254 32 : if (non_null_cnt > 0)
255 : {
256 : Datum *bound_hist_values;
257 : Datum *length_hist_values;
258 : int pos,
259 : posfrac,
260 : delta,
261 : deltafrac,
262 : i;
263 : MemoryContext old_cxt;
264 : float4 *emptyfrac;
265 :
266 32 : stats->stats_valid = true;
267 : /* Do the simple null-frac and width stats */
268 32 : stats->stanullfrac = (double) null_cnt / (double) samplerows;
269 32 : stats->stawidth = total_width / (double) non_null_cnt;
270 :
271 : /* Estimate that non-null values are unique */
272 32 : stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac);
273 :
274 : /* Must copy the target values into anl_context */
275 32 : old_cxt = MemoryContextSwitchTo(stats->anl_context);
276 :
277 : /*
278 : * Generate a bounds histogram slot entry if there are at least two
279 : * values.
280 : */
281 32 : if (non_empty_cnt >= 2)
282 : {
283 : /* Sort bound values */
284 32 : qsort_interruptible(lowers, non_empty_cnt, sizeof(RangeBound),
285 : range_bound_qsort_cmp, typcache);
286 32 : qsort_interruptible(uppers, non_empty_cnt, sizeof(RangeBound),
287 : range_bound_qsort_cmp, typcache);
288 :
289 32 : num_hist = non_empty_cnt;
290 32 : if (num_hist > num_bins)
291 20 : num_hist = num_bins + 1;
292 :
293 32 : bound_hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
294 :
295 : /*
296 : * The object of this loop is to construct ranges from first and
297 : * last entries in lowers[] and uppers[] along with evenly-spaced
298 : * values in between. So the i'th value is a range of lowers[(i *
299 : * (nvals - 1)) / (num_hist - 1)] and uppers[(i * (nvals - 1)) /
300 : * (num_hist - 1)]. But computing that subscript directly risks
301 : * integer overflow when the stats target is more than a couple
302 : * thousand. Instead we add (nvals - 1) / (num_hist - 1) to pos
303 : * at each step, tracking the integral and fractional parts of the
304 : * sum separately.
305 : */
306 32 : delta = (non_empty_cnt - 1) / (num_hist - 1);
307 32 : deltafrac = (non_empty_cnt - 1) % (num_hist - 1);
308 32 : pos = posfrac = 0;
309 :
310 2124 : for (i = 0; i < num_hist; i++)
311 : {
312 2092 : bound_hist_values[i] = PointerGetDatum(range_serialize(typcache,
313 2092 : &lowers[pos],
314 2092 : &uppers[pos],
315 : false,
316 : NULL));
317 2092 : pos += delta;
318 2092 : posfrac += deltafrac;
319 2092 : if (posfrac >= (num_hist - 1))
320 : {
321 : /* fractional part exceeds 1, carry to integer part */
322 1980 : pos++;
323 1980 : posfrac -= (num_hist - 1);
324 : }
325 : }
326 :
327 32 : stats->stakind[slot_idx] = STATISTIC_KIND_BOUNDS_HISTOGRAM;
328 32 : stats->stavalues[slot_idx] = bound_hist_values;
329 32 : stats->numvalues[slot_idx] = num_hist;
330 :
331 : /* Store ranges even if we're analyzing a multirange column */
332 32 : stats->statypid[slot_idx] = typcache->type_id;
333 32 : stats->statyplen[slot_idx] = typcache->typlen;
334 32 : stats->statypbyval[slot_idx] = typcache->typbyval;
335 32 : stats->statypalign[slot_idx] = typcache->typalign;
336 :
337 32 : slot_idx++;
338 : }
339 :
340 : /*
341 : * Generate a length histogram slot entry if there are at least two
342 : * values.
343 : */
344 32 : if (non_empty_cnt >= 2)
345 : {
346 : /*
347 : * Ascending sort of range lengths for further filling of
348 : * histogram
349 : */
350 32 : qsort_interruptible(lengths, non_empty_cnt, sizeof(float8),
351 : float8_qsort_cmp, NULL);
352 :
353 32 : num_hist = non_empty_cnt;
354 32 : if (num_hist > num_bins)
355 20 : num_hist = num_bins + 1;
356 :
357 32 : length_hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
358 :
359 : /*
360 : * The object of this loop is to copy the first and last lengths[]
361 : * entries along with evenly-spaced values in between. So the i'th
362 : * value is lengths[(i * (nvals - 1)) / (num_hist - 1)]. But
363 : * computing that subscript directly risks integer overflow when
364 : * the stats target is more than a couple thousand. Instead we
365 : * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
366 : * the integral and fractional parts of the sum separately.
367 : */
368 32 : delta = (non_empty_cnt - 1) / (num_hist - 1);
369 32 : deltafrac = (non_empty_cnt - 1) % (num_hist - 1);
370 32 : pos = posfrac = 0;
371 :
372 2124 : for (i = 0; i < num_hist; i++)
373 : {
374 2092 : length_hist_values[i] = Float8GetDatum(lengths[pos]);
375 2092 : pos += delta;
376 2092 : posfrac += deltafrac;
377 2092 : if (posfrac >= (num_hist - 1))
378 : {
379 : /* fractional part exceeds 1, carry to integer part */
380 1980 : pos++;
381 1980 : posfrac -= (num_hist - 1);
382 : }
383 : }
384 : }
385 : else
386 : {
387 : /*
388 : * Even when we don't create the histogram, store an empty array
389 : * to mean "no histogram". We can't just leave stavalues NULL,
390 : * because get_attstatsslot() errors if you ask for stavalues, and
391 : * it's NULL. We'll still store the empty fraction in stanumbers.
392 : */
393 0 : length_hist_values = palloc(0);
394 0 : num_hist = 0;
395 : }
396 32 : stats->staop[slot_idx] = Float8LessOperator;
397 32 : stats->stacoll[slot_idx] = InvalidOid;
398 32 : stats->stavalues[slot_idx] = length_hist_values;
399 32 : stats->numvalues[slot_idx] = num_hist;
400 32 : stats->statypid[slot_idx] = FLOAT8OID;
401 32 : stats->statyplen[slot_idx] = sizeof(float8);
402 32 : stats->statypbyval[slot_idx] = FLOAT8PASSBYVAL;
403 32 : stats->statypalign[slot_idx] = 'd';
404 :
405 : /* Store the fraction of empty ranges */
406 32 : emptyfrac = (float4 *) palloc(sizeof(float4));
407 32 : *emptyfrac = ((double) empty_cnt) / ((double) non_null_cnt);
408 32 : stats->stanumbers[slot_idx] = emptyfrac;
409 32 : stats->numnumbers[slot_idx] = 1;
410 :
411 32 : stats->stakind[slot_idx] = STATISTIC_KIND_RANGE_LENGTH_HISTOGRAM;
412 32 : slot_idx++;
413 :
414 32 : MemoryContextSwitchTo(old_cxt);
415 : }
416 0 : else if (null_cnt > 0)
417 : {
418 : /* We found only nulls; assume the column is entirely null */
419 0 : stats->stats_valid = true;
420 0 : stats->stanullfrac = 1.0;
421 0 : stats->stawidth = 0; /* "unknown" */
422 0 : stats->stadistinct = 0.0; /* "unknown" */
423 : }
424 :
425 : /*
426 : * We don't need to bother cleaning up any of our temporary palloc's. The
427 : * hashtable should also go away, as it used a child memory context.
428 : */
429 32 : }
|
