PostGIS  2.5.0beta2dev-r@@SVN_REVISION@@

◆ estimate_selectivity()

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 1913 of file gserialized_estimate.c.

References _postgis_gserialized_stats(), 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, PG_FUNCTION_INFO_V1(), ND_STATS_T::size, and ND_STATS_T::value.

Referenced by _postgis_gserialized_sel(), and gserialized_gist_sel().

1914 {
1915  int d; /* counter */
1916  float8 selectivity;
1917  ND_BOX nd_box;
1918  ND_IBOX nd_ibox;
1919  int at[ND_DIMS];
1920  double cell_size[ND_DIMS];
1921  double min[ND_DIMS];
1922  double max[ND_DIMS];
1923  double total_count = 0.0;
1924  int ndims_max;
1925 
1926  /* Calculate the overlap of the box on the histogram */
1927  if ( ! nd_stats )
1928  {
1929  elog(NOTICE, " estimate_selectivity called with null input");
1930  return FALLBACK_ND_SEL;
1931  }
1932 
1933  ndims_max = Max(nd_stats->ndims, gbox_ndims(box));
1934 
1935  /* Initialize nd_box. */
1936  nd_box_from_gbox(box, &nd_box);
1937 
1938  /*
1939  * To return 2D stats on an ND sample, we need to make the
1940  * 2D box cover the full range of the other dimensions in the
1941  * histogram.
1942  */
1943  POSTGIS_DEBUGF(3, " mode: %d", mode);
1944  if ( mode == 2 )
1945  {
1946  POSTGIS_DEBUG(3, " in 2d mode, stripping the computation down to 2d");
1947  ndims_max = 2;
1948  }
1949 
1950  POSTGIS_DEBUGF(3, " nd_stats->extent: %s", nd_box_to_json(&(nd_stats->extent), nd_stats->ndims));
1951  POSTGIS_DEBUGF(3, " nd_box: %s", nd_box_to_json(&(nd_box), gbox_ndims(box)));
1952 
1953  /*
1954  * Search box completely misses histogram extent?
1955  * We have to intersect in all N dimensions or else we have
1956  * zero interaction under the &&& operator. It's important
1957  * to short circuit in this case, as some of the tests below
1958  * will return junk results when run on non-intersecting inputs.
1959  */
1960  if ( ! nd_box_intersects(&nd_box, &(nd_stats->extent), ndims_max) )
1961  {
1962  POSTGIS_DEBUG(3, " search box does not overlap histogram, returning 0");
1963  return 0.0;
1964  }
1965 
1966  /* Search box completely contains histogram extent! */
1967  if ( nd_box_contains(&nd_box, &(nd_stats->extent), ndims_max) )
1968  {
1969  POSTGIS_DEBUG(3, " search box contains histogram, returning 1");
1970  return 1.0;
1971  }
1972 
1973  /* Calculate the overlap of the box on the histogram */
1974  if ( ! nd_box_overlap(nd_stats, &nd_box, &nd_ibox) )
1975  {
1976  POSTGIS_DEBUG(3, " search box overlap with stats histogram failed");
1977  return FALLBACK_ND_SEL;
1978  }
1979 
1980  /* Work out some measurements of the histogram */
1981  for ( d = 0; d < nd_stats->ndims; d++ )
1982  {
1983  /* Cell size in each dim */
1984  min[d] = nd_stats->extent.min[d];
1985  max[d] = nd_stats->extent.max[d];
1986  cell_size[d] = (max[d] - min[d]) / nd_stats->size[d];
1987  POSTGIS_DEBUGF(3, " cell_size[%d] : %.9g", d, cell_size[d]);
1988 
1989  /* Initialize the counter */
1990  at[d] = nd_ibox.min[d];
1991  }
1992 
1993  /* Move through all the overlap values and sum them */
1994  do
1995  {
1996  float cell_count, ratio;
1997  ND_BOX nd_cell = { {0.0, 0.0, 0.0, 0.0}, {0.0, 0.0, 0.0, 0.0} };
1998 
1999  /* We have to pro-rate partially overlapped cells. */
2000  for ( d = 0; d < nd_stats->ndims; d++ )
2001  {
2002  nd_cell.min[d] = min[d] + (at[d]+0) * cell_size[d];
2003  nd_cell.max[d] = min[d] + (at[d]+1) * cell_size[d];
2004  }
2005 
2006  ratio = nd_box_ratio(&nd_box, &nd_cell, nd_stats->ndims);
2007  cell_count = nd_stats->value[nd_stats_value_index(nd_stats, at)];
2008 
2009  /* Add the pro-rated count for this cell to the overall total */
2010  total_count += cell_count * ratio;
2011  POSTGIS_DEBUGF(4, " cell (%d,%d), cell value %.6f, ratio %.6f", at[0], at[1], cell_count, ratio);
2012  }
2013  while ( nd_increment(&nd_ibox, nd_stats->ndims, at) );
2014 
2015  /* Scale by the number of features in our histogram to get the proportion */
2016  selectivity = total_count / nd_stats->histogram_features;
2017 
2018  POSTGIS_DEBUGF(3, " nd_stats->histogram_features = %f", nd_stats->histogram_features);
2019  POSTGIS_DEBUGF(3, " nd_stats->histogram_cells = %f", nd_stats->histogram_cells);
2020  POSTGIS_DEBUGF(3, " sum(overlapped histogram cells) = %f", total_count);
2021  POSTGIS_DEBUGF(3, " selectivity = %f", selectivity);
2022 
2023  /* Prevent rounding overflows */
2024  if (selectivity > 1.0) selectivity = 1.0;
2025  else if (selectivity < 0.0) selectivity = 0.0;
2026 
2027  return selectivity;
2028 }
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: