PostGIS  2.3.7dev-r@@SVN_REVISION@@
static float8 estimate_selectivity ( const GBOX box,
const ND_STATS nd_stats,
int  mode 
)
static

This function returns an estimate of the selectivity of a search GBOX by looking at data in the ND_STATS structure.

The selectivity is a float from 0-1, that estimates the proportion of the rows in the table that will be returned as a result of the search box.

To get our estimate, we need "only" sum up the values * the proportion of each cell in the histogram that falls within the search box, then divide by the number of features that generated the histogram.

Definition at line 1864 of file gserialized_estimate.c.

References ND_STATS_T::extent, FALLBACK_ND_SEL, gbox_ndims(), ND_STATS_T::histogram_cells, ND_STATS_T::histogram_features, ND_BOX_T::max, ND_BOX_T::min, ND_IBOX_T::min, nd_box_contains(), nd_box_from_gbox(), nd_box_intersects(), nd_box_overlap(), nd_box_ratio(), nd_box_to_json(), ND_DIMS, nd_increment(), nd_stats_value_index(), ND_STATS_T::ndims, ND_STATS_T::size, and ND_STATS_T::value.

Referenced by _postgis_gserialized_sel(), and gserialized_gist_sel().

1865 {
1866  int d; /* counter */
1867  float8 selectivity;
1868  ND_BOX nd_box;
1869  ND_IBOX nd_ibox;
1870  int at[ND_DIMS];
1871  double cell_size[ND_DIMS];
1872  double min[ND_DIMS];
1873  double max[ND_DIMS];
1874  double total_count = 0.0;
1875  int ndims_max = Max(nd_stats->ndims, gbox_ndims(box));
1876 // int ndims_min = Min(nd_stats->ndims, gbox_ndims(box));
1877 
1878  /* Calculate the overlap of the box on the histogram */
1879  if ( ! nd_stats )
1880  {
1881  elog(NOTICE, " estimate_selectivity called with null input");
1882  return FALLBACK_ND_SEL;
1883  }
1884 
1885  /* Initialize nd_box. */
1886  nd_box_from_gbox(box, &nd_box);
1887 
1888  /*
1889  * To return 2D stats on an ND sample, we need to make the
1890  * 2D box cover the full range of the other dimensions in the
1891  * histogram.
1892  */
1893  POSTGIS_DEBUGF(3, " mode: %d", mode);
1894  if ( mode == 2 )
1895  {
1896  POSTGIS_DEBUG(3, " in 2d mode, stripping the computation down to 2d");
1897  ndims_max = 2;
1898  }
1899 
1900  POSTGIS_DEBUGF(3, " nd_stats->extent: %s", nd_box_to_json(&(nd_stats->extent), nd_stats->ndims));
1901  POSTGIS_DEBUGF(3, " nd_box: %s", nd_box_to_json(&(nd_box), gbox_ndims(box)));
1902 
1903  /*
1904  * Search box completely misses histogram extent?
1905  * We have to intersect in all N dimensions or else we have
1906  * zero interaction under the &&& operator. It's important
1907  * to short circuit in this case, as some of the tests below
1908  * will return junk results when run on non-intersecting inputs.
1909  */
1910  if ( ! nd_box_intersects(&nd_box, &(nd_stats->extent), ndims_max) )
1911  {
1912  POSTGIS_DEBUG(3, " search box does not overlap histogram, returning 0");
1913  return 0.0;
1914  }
1915 
1916  /* Search box completely contains histogram extent! */
1917  if ( nd_box_contains(&nd_box, &(nd_stats->extent), ndims_max) )
1918  {
1919  POSTGIS_DEBUG(3, " search box contains histogram, returning 1");
1920  return 1.0;
1921  }
1922 
1923  /* Calculate the overlap of the box on the histogram */
1924  if ( ! nd_box_overlap(nd_stats, &nd_box, &nd_ibox) )
1925  {
1926  POSTGIS_DEBUG(3, " search box overlap with stats histogram failed");
1927  return FALLBACK_ND_SEL;
1928  }
1929 
1930  /* Work out some measurements of the histogram */
1931  for ( d = 0; d < nd_stats->ndims; d++ )
1932  {
1933  /* Cell size in each dim */
1934  min[d] = nd_stats->extent.min[d];
1935  max[d] = nd_stats->extent.max[d];
1936  cell_size[d] = (max[d] - min[d]) / nd_stats->size[d];
1937  POSTGIS_DEBUGF(3, " cell_size[%d] : %.9g", d, cell_size[d]);
1938 
1939  /* Initialize the counter */
1940  at[d] = nd_ibox.min[d];
1941  }
1942 
1943  /* Move through all the overlap values and sum them */
1944  do
1945  {
1946  float cell_count, ratio;
1947  ND_BOX nd_cell;
1948 
1949  /* We have to pro-rate partially overlapped cells. */
1950  for ( d = 0; d < nd_stats->ndims; d++ )
1951  {
1952  nd_cell.min[d] = min[d] + (at[d]+0) * cell_size[d];
1953  nd_cell.max[d] = min[d] + (at[d]+1) * cell_size[d];
1954  }
1955 
1956  ratio = nd_box_ratio(&nd_box, &nd_cell, nd_stats->ndims);
1957  cell_count = nd_stats->value[nd_stats_value_index(nd_stats, at)];
1958 
1959  /* Add the pro-rated count for this cell to the overall total */
1960  total_count += cell_count * ratio;
1961  POSTGIS_DEBUGF(4, " cell (%d,%d), cell value %.6f, ratio %.6f", at[0], at[1], cell_count, ratio);
1962  }
1963  while ( nd_increment(&nd_ibox, nd_stats->ndims, at) );
1964 
1965  /* Scale by the number of features in our histogram to get the proportion */
1966  selectivity = total_count / nd_stats->histogram_features;
1967 
1968  POSTGIS_DEBUGF(3, " nd_stats->histogram_features = %f", nd_stats->histogram_features);
1969  POSTGIS_DEBUGF(3, " nd_stats->histogram_cells = %f", nd_stats->histogram_cells);
1970  POSTGIS_DEBUGF(3, " sum(overlapped histogram cells) = %f", total_count);
1971  POSTGIS_DEBUGF(3, " selectivity = %f", selectivity);
1972 
1973  /* Prevent rounding overflows */
1974  if (selectivity > 1.0) selectivity = 1.0;
1975  else if (selectivity < 0.0) selectivity = 0.0;
1976 
1977  return selectivity;
1978 }
static int nd_increment(ND_IBOX *ibox, int ndims, int *counter)
Given an n-d index array (counter), and a domain to increment it in (ibox) increment it by one...
static int nd_box_overlap(const ND_STATS *nd_stats, const ND_BOX *nd_box, ND_IBOX *nd_ibox)
What stats cells overlap with this ND_BOX? Put the lowest cell addresses in ND_IBOX->min and the high...
#define ND_DIMS
The maximum number of dimensions our code can handle.
static int nd_box_contains(const ND_BOX *a, const ND_BOX *b, int ndims)
Return TRUE if ND_BOX a contains b, false otherwise.
static void nd_box_from_gbox(const GBOX *gbox, ND_BOX *nd_box)
Set the values of an ND_BOX from a GBOX.
#define FALLBACK_ND_SEL
More modest fallback selectivity factor.
static int nd_stats_value_index(const ND_STATS *stats, int *indexes)
Given a position in the n-d histogram (i,j,k) return the position in the 1-d values array...
float4 size[ND_DIMS]
static int nd_box_intersects(const ND_BOX *a, const ND_BOX *b, int ndims)
Return TRUE if ND_BOX a overlaps b, false otherwise.
int min[ND_DIMS]
N-dimensional box index type.
static char * nd_box_to_json(const ND_BOX *nd_box, int ndims)
Convert an ND_BOX to a JSON string for printing.
float4 max[ND_DIMS]
float4 min[ND_DIMS]
static double nd_box_ratio(const ND_BOX *b1, const ND_BOX *b2, int ndims)
Returns the proportion of b2 that is covered by b1.
N-dimensional box type for calculations, to avoid doing explicit axis conversions from GBOX in all ca...
static int gbox_ndims(const GBOX *gbox)
Given that geodetic boxes are X/Y/Z regardless of the underlying geometry dimensionality and other bo...

Here is the call graph for this function:

Here is the caller graph for this function: