1 | /*------------------------------------------------------------------------- |
2 | * |
3 | * ragetypes_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) 1996-2019, 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 | |
34 | static int float8_qsort_cmp(const void *a1, const void *a2); |
35 | static int range_bound_qsort_cmp(const void *a1, const void *a2, void *arg); |
36 | static void compute_range_stats(VacAttrStats *stats, |
37 | AnalyzeAttrFetchFunc fetchfunc, int samplerows, double totalrows); |
38 | |
39 | /* |
40 | * range_typanalyze -- typanalyze function for range columns |
41 | */ |
42 | Datum |
43 | range_typanalyze(PG_FUNCTION_ARGS) |
44 | { |
45 | VacAttrStats *stats = (VacAttrStats *) PG_GETARG_POINTER(0); |
46 | TypeCacheEntry *typcache; |
47 | Form_pg_attribute attr = stats->attr; |
48 | |
49 | /* Get information about range type; note column might be a domain */ |
50 | typcache = range_get_typcache(fcinfo, getBaseType(stats->attrtypid)); |
51 | |
52 | if (attr->attstattarget < 0) |
53 | attr->attstattarget = default_statistics_target; |
54 | |
55 | stats->compute_stats = compute_range_stats; |
56 | stats->extra_data = typcache; |
57 | /* same as in std_typanalyze */ |
58 | stats->minrows = 300 * attr->attstattarget; |
59 | |
60 | PG_RETURN_BOOL(true); |
61 | } |
62 | |
63 | /* |
64 | * Comparison function for sorting float8s, used for range lengths. |
65 | */ |
66 | static int |
67 | float8_qsort_cmp(const void *a1, const void *a2) |
68 | { |
69 | const float8 *f1 = (const float8 *) a1; |
70 | const float8 *f2 = (const float8 *) a2; |
71 | |
72 | if (*f1 < *f2) |
73 | return -1; |
74 | else if (*f1 == *f2) |
75 | return 0; |
76 | else |
77 | return 1; |
78 | } |
79 | |
80 | /* |
81 | * Comparison function for sorting RangeBounds. |
82 | */ |
83 | static int |
84 | range_bound_qsort_cmp(const void *a1, const void *a2, void *arg) |
85 | { |
86 | RangeBound *b1 = (RangeBound *) a1; |
87 | RangeBound *b2 = (RangeBound *) a2; |
88 | TypeCacheEntry *typcache = (TypeCacheEntry *) arg; |
89 | |
90 | return range_cmp_bounds(typcache, b1, b2); |
91 | } |
92 | |
93 | /* |
94 | * compute_range_stats() -- compute statistics for a range column |
95 | */ |
96 | static void |
97 | compute_range_stats(VacAttrStats *stats, AnalyzeAttrFetchFunc fetchfunc, |
98 | int samplerows, double totalrows) |
99 | { |
100 | TypeCacheEntry *typcache = (TypeCacheEntry *) stats->extra_data; |
101 | bool has_subdiff = OidIsValid(typcache->rng_subdiff_finfo.fn_oid); |
102 | int null_cnt = 0; |
103 | int non_null_cnt = 0; |
104 | int non_empty_cnt = 0; |
105 | int empty_cnt = 0; |
106 | int range_no; |
107 | int slot_idx; |
108 | int num_bins = stats->attr->attstattarget; |
109 | int num_hist; |
110 | float8 *lengths; |
111 | RangeBound *lowers, |
112 | *uppers; |
113 | double total_width = 0; |
114 | |
115 | /* Allocate memory to hold range bounds and lengths of the sample ranges. */ |
116 | lowers = (RangeBound *) palloc(sizeof(RangeBound) * samplerows); |
117 | uppers = (RangeBound *) palloc(sizeof(RangeBound) * samplerows); |
118 | lengths = (float8 *) palloc(sizeof(float8) * samplerows); |
119 | |
120 | /* Loop over the sample ranges. */ |
121 | for (range_no = 0; range_no < samplerows; range_no++) |
122 | { |
123 | Datum value; |
124 | bool isnull, |
125 | empty; |
126 | RangeType *range; |
127 | RangeBound lower, |
128 | upper; |
129 | float8 length; |
130 | |
131 | vacuum_delay_point(); |
132 | |
133 | value = fetchfunc(stats, range_no, &isnull); |
134 | if (isnull) |
135 | { |
136 | /* range is null, just count that */ |
137 | null_cnt++; |
138 | continue; |
139 | } |
140 | |
141 | /* |
142 | * XXX: should we ignore wide values, like std_typanalyze does, to |
143 | * avoid bloating the statistics table? |
144 | */ |
145 | total_width += VARSIZE_ANY(DatumGetPointer(value)); |
146 | |
147 | /* Get range and deserialize it for further analysis. */ |
148 | range = DatumGetRangeTypeP(value); |
149 | range_deserialize(typcache, range, &lower, &upper, &empty); |
150 | |
151 | if (!empty) |
152 | { |
153 | /* Remember bounds and length for further usage in histograms */ |
154 | lowers[non_empty_cnt] = lower; |
155 | uppers[non_empty_cnt] = upper; |
156 | |
157 | if (lower.infinite || upper.infinite) |
158 | { |
159 | /* Length of any kind of an infinite range is infinite */ |
160 | length = get_float8_infinity(); |
161 | } |
162 | else if (has_subdiff) |
163 | { |
164 | /* |
165 | * For an ordinary range, use subdiff function between upper |
166 | * and lower bound values. |
167 | */ |
168 | length = DatumGetFloat8(FunctionCall2Coll( |
169 | &typcache->rng_subdiff_finfo, |
170 | typcache->rng_collation, |
171 | upper.val, lower.val)); |
172 | } |
173 | else |
174 | { |
175 | /* Use default value of 1.0 if no subdiff is available. */ |
176 | length = 1.0; |
177 | } |
178 | lengths[non_empty_cnt] = length; |
179 | |
180 | non_empty_cnt++; |
181 | } |
182 | else |
183 | empty_cnt++; |
184 | |
185 | non_null_cnt++; |
186 | } |
187 | |
188 | slot_idx = 0; |
189 | |
190 | /* We can only compute real stats if we found some non-null values. */ |
191 | if (non_null_cnt > 0) |
192 | { |
193 | Datum *bound_hist_values; |
194 | Datum *length_hist_values; |
195 | int pos, |
196 | posfrac, |
197 | delta, |
198 | deltafrac, |
199 | i; |
200 | MemoryContext old_cxt; |
201 | float4 *emptyfrac; |
202 | |
203 | stats->stats_valid = true; |
204 | /* Do the simple null-frac and width stats */ |
205 | stats->stanullfrac = (double) null_cnt / (double) samplerows; |
206 | stats->stawidth = total_width / (double) non_null_cnt; |
207 | |
208 | /* Estimate that non-null values are unique */ |
209 | stats->stadistinct = -1.0 * (1.0 - stats->stanullfrac); |
210 | |
211 | /* Must copy the target values into anl_context */ |
212 | old_cxt = MemoryContextSwitchTo(stats->anl_context); |
213 | |
214 | /* |
215 | * Generate a bounds histogram slot entry if there are at least two |
216 | * values. |
217 | */ |
218 | if (non_empty_cnt >= 2) |
219 | { |
220 | /* Sort bound values */ |
221 | qsort_arg(lowers, non_empty_cnt, sizeof(RangeBound), |
222 | range_bound_qsort_cmp, typcache); |
223 | qsort_arg(uppers, non_empty_cnt, sizeof(RangeBound), |
224 | range_bound_qsort_cmp, typcache); |
225 | |
226 | num_hist = non_empty_cnt; |
227 | if (num_hist > num_bins) |
228 | num_hist = num_bins + 1; |
229 | |
230 | bound_hist_values = (Datum *) palloc(num_hist * sizeof(Datum)); |
231 | |
232 | /* |
233 | * The object of this loop is to construct ranges from first and |
234 | * last entries in lowers[] and uppers[] along with evenly-spaced |
235 | * values in between. So the i'th value is a range of lowers[(i * |
236 | * (nvals - 1)) / (num_hist - 1)] and uppers[(i * (nvals - 1)) / |
237 | * (num_hist - 1)]. But computing that subscript directly risks |
238 | * integer overflow when the stats target is more than a couple |
239 | * thousand. Instead we add (nvals - 1) / (num_hist - 1) to pos |
240 | * at each step, tracking the integral and fractional parts of the |
241 | * sum separately. |
242 | */ |
243 | delta = (non_empty_cnt - 1) / (num_hist - 1); |
244 | deltafrac = (non_empty_cnt - 1) % (num_hist - 1); |
245 | pos = posfrac = 0; |
246 | |
247 | for (i = 0; i < num_hist; i++) |
248 | { |
249 | bound_hist_values[i] = PointerGetDatum(range_serialize( |
250 | typcache, &lowers[pos], &uppers[pos], false)); |
251 | pos += delta; |
252 | posfrac += deltafrac; |
253 | if (posfrac >= (num_hist - 1)) |
254 | { |
255 | /* fractional part exceeds 1, carry to integer part */ |
256 | pos++; |
257 | posfrac -= (num_hist - 1); |
258 | } |
259 | } |
260 | |
261 | stats->stakind[slot_idx] = STATISTIC_KIND_BOUNDS_HISTOGRAM; |
262 | stats->stavalues[slot_idx] = bound_hist_values; |
263 | stats->numvalues[slot_idx] = num_hist; |
264 | slot_idx++; |
265 | } |
266 | |
267 | /* |
268 | * Generate a length histogram slot entry if there are at least two |
269 | * values. |
270 | */ |
271 | if (non_empty_cnt >= 2) |
272 | { |
273 | /* |
274 | * Ascending sort of range lengths for further filling of |
275 | * histogram |
276 | */ |
277 | qsort(lengths, non_empty_cnt, sizeof(float8), float8_qsort_cmp); |
278 | |
279 | num_hist = non_empty_cnt; |
280 | if (num_hist > num_bins) |
281 | num_hist = num_bins + 1; |
282 | |
283 | length_hist_values = (Datum *) palloc(num_hist * sizeof(Datum)); |
284 | |
285 | /* |
286 | * The object of this loop is to copy the first and last lengths[] |
287 | * entries along with evenly-spaced values in between. So the i'th |
288 | * value is lengths[(i * (nvals - 1)) / (num_hist - 1)]. But |
289 | * computing that subscript directly risks integer overflow when |
290 | * the stats target is more than a couple thousand. Instead we |
291 | * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking |
292 | * the integral and fractional parts of the sum separately. |
293 | */ |
294 | delta = (non_empty_cnt - 1) / (num_hist - 1); |
295 | deltafrac = (non_empty_cnt - 1) % (num_hist - 1); |
296 | pos = posfrac = 0; |
297 | |
298 | for (i = 0; i < num_hist; i++) |
299 | { |
300 | length_hist_values[i] = Float8GetDatum(lengths[pos]); |
301 | pos += delta; |
302 | posfrac += deltafrac; |
303 | if (posfrac >= (num_hist - 1)) |
304 | { |
305 | /* fractional part exceeds 1, carry to integer part */ |
306 | pos++; |
307 | posfrac -= (num_hist - 1); |
308 | } |
309 | } |
310 | } |
311 | else |
312 | { |
313 | /* |
314 | * Even when we don't create the histogram, store an empty array |
315 | * to mean "no histogram". We can't just leave stavalues NULL, |
316 | * because get_attstatsslot() errors if you ask for stavalues, and |
317 | * it's NULL. We'll still store the empty fraction in stanumbers. |
318 | */ |
319 | length_hist_values = palloc(0); |
320 | num_hist = 0; |
321 | } |
322 | stats->staop[slot_idx] = Float8LessOperator; |
323 | stats->stacoll[slot_idx] = InvalidOid; |
324 | stats->stavalues[slot_idx] = length_hist_values; |
325 | stats->numvalues[slot_idx] = num_hist; |
326 | stats->statypid[slot_idx] = FLOAT8OID; |
327 | stats->statyplen[slot_idx] = sizeof(float8); |
328 | #ifdef USE_FLOAT8_BYVAL |
329 | stats->statypbyval[slot_idx] = true; |
330 | #else |
331 | stats->statypbyval[slot_idx] = false; |
332 | #endif |
333 | stats->statypalign[slot_idx] = 'd'; |
334 | |
335 | /* Store the fraction of empty ranges */ |
336 | emptyfrac = (float4 *) palloc(sizeof(float4)); |
337 | *emptyfrac = ((double) empty_cnt) / ((double) non_null_cnt); |
338 | stats->stanumbers[slot_idx] = emptyfrac; |
339 | stats->numnumbers[slot_idx] = 1; |
340 | |
341 | stats->stakind[slot_idx] = STATISTIC_KIND_RANGE_LENGTH_HISTOGRAM; |
342 | slot_idx++; |
343 | |
344 | MemoryContextSwitchTo(old_cxt); |
345 | } |
346 | else if (null_cnt > 0) |
347 | { |
348 | /* We found only nulls; assume the column is entirely null */ |
349 | stats->stats_valid = true; |
350 | stats->stanullfrac = 1.0; |
351 | stats->stawidth = 0; /* "unknown" */ |
352 | stats->stadistinct = 0.0; /* "unknown" */ |
353 | } |
354 | |
355 | /* |
356 | * We don't need to bother cleaning up any of our temporary palloc's. The |
357 | * hashtable should also go away, as it used a child memory context. |
358 | */ |
359 | } |
360 | |